INFO 04-12 10:44:11 [__init__.py:216] Automatically detected platform cuda.
2026-04-12 10:44:12,700 | INFO | train_grpo_train | mode=train

Fetching 128 files:   0%|          | 0/128 [00:00<?, ?it/s]
Fetching 128 files: 100%|██████████| 128/128 [00:00<00:00, 12265.17it/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,  1.34it/s]
Loading checkpoint shards:  67%|██████▋   | 2/3 [00:01<00:00,  1.35it/s]
Loading checkpoint shards: 100%|██████████| 3/3 [00:02<00:00,  1.41it/s]
Loading checkpoint shards: 100%|██████████| 3/3 [00:02<00:00,  1.39it/s]
The model is already on multiple devices. Skipping the move to device specified in `args`.
INFO 04-12 10:45:52 [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 04-12 10:45:53 [model.py:547] Resolved architecture: LlamaForCausalLM
INFO 04-12 10:45:53 [model.py:1510] Using max model len 1536
INFO 04-12 10:45:53 [parallel.py:380] Using external launcher for distributed inference.
INFO 04-12 10:45:53 [parallel.py:420] Disabling V1 multiprocessing for external launcher.
INFO 04-12 10:45:53 [scheduler.py:205] Chunked prefill is enabled with max_num_batched_tokens=4096.
INFO 04-12 10:45:53 [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 04-12 10:45:54 [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 04-12 10:45:54 [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 04-12 10:45:55 [gpu_model_runner.py:2602] Starting to load model Navid-AI/Yehia-7B-preview...
INFO 04-12 10:45:55 [gpu_model_runner.py:2634] Loading model from scratch...
INFO 04-12 10:45:55 [cuda.py:366] Using Flash Attention backend on V1 engine.
INFO 04-12 10:45:55 [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<00:00,  2.11it/s]

Loading safetensors checkpoint shards:  67% Completed | 2/3 [00:01<00:00,  1.72it/s]

Loading safetensors checkpoint shards: 100% Completed | 3/3 [00:01<00:00,  1.55it/s]

Loading safetensors checkpoint shards: 100% Completed | 3/3 [00:01<00:00,  1.62it/s]

INFO 04-12 10:45:57 [default_loader.py:267] Loading weights took 2.06 seconds
INFO 04-12 10:45:58 [gpu_model_runner.py:2653] Model loading took 13.0406 GiB and 2.502332 seconds
INFO 04-12 10:46:03 [backends.py:548] Using cache directory: /root/.cache/vllm/torch_compile_cache/8b012be8b8/rank_0_0/backbone for vLLM's torch.compile
INFO 04-12 10:46:03 [backends.py:559] Dynamo bytecode transform time: 5.04 s
INFO 04-12 10:46:07 [backends.py:164] Directly load the compiled graph(s) for dynamic shape from the cache, took 3.678 s
INFO 04-12 10:46:08 [monitor.py:34] torch.compile takes 5.04 s in total
INFO 04-12 10:46:09 [gpu_worker.py:298] Available KV cache memory: 26.11 GiB
INFO 04-12 10:46:10 [kv_cache_utils.py:1087] GPU KV cache size: 53,456 tokens
INFO 04-12 10:46:10 [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, 19.14it/s]
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE): 100%|██████████| 5/5 [00:00<00:00, 17.70it/s]
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE): 100%|██████████| 5/5 [00:00<00:00, 17.85it/s]

Capturing CUDA graphs (decode, FULL):   0%|          | 0/4 [00:00<?, ?it/s]
Capturing CUDA graphs (decode, FULL):  75%|███████▌  | 3/4 [00:00<00:00, 23.48it/s]
Capturing CUDA graphs (decode, FULL): 100%|██████████| 4/4 [00:00<00:00, 23.74it/s]
INFO 04-12 10:46:11 [gpu_model_runner.py:3480] Graph capturing finished in 1 secs, took 0.09 GiB
INFO 04-12 10:46:11 [core.py:210] init engine (profile, create kv cache, warmup model) took 12.80 seconds
INFO 04-12 10:46:11 [llm.py:306] Supported_tasks: ('generate',)
2026-04-12 10:46:11,585 | 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
2026-04-12 10:46:11,585 | 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
2026-04-12 10:46:11,585 | INFO | train_grpo_train | active_rewards=['meter', 'count_adherence', 'arabic_clean', 'lexical_plausibility', 'judge_quality', 'repeat_penalty', 'near_duplicate_penalty', 'opening_diversity', 'distinct_2', 'total_composite']
2026-04-12 10:46:11,585 | INFO | train_grpo_train | reward_weights={'meter': 0.0, 'count_adherence': 0.0, 'arabic_clean': 0.0, 'lexical_plausibility': 0.0, 'judge_quality': 0.0, 'repeat_penalty': 0.0, 'near_duplicate_penalty': 0.0, 'opening_diversity': 0.0, 'distinct_2': 0.0, 'total_composite': 1.0}
2026-04-12 10:46:11,585 | INFO | train_grpo_train | output_repo=Shaer-AI/Shaer-adapters-grpo-vnext
2026-04-12 10:46:11,585 | INFO | train_grpo_train | train_manifest_path=/root/workspace/Shaer/grpo/outputs/curated_meter_count_drop_trio/cap_3000/selected_manifest.csv
2026-04-12 10:46:11,585 | 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
2026-04-12 10:46:11,585 | INFO | train_grpo_train | trainable_params=159907840 total_params=7160467456 trainable_ratio=0.022332
2026-04-12 10:46:11,586 | INFO | train_grpo_train | resume_decision={"timestamp_utc": "2026-04-12T10:45:52Z", "resume_mode": "fresh", "requested_spec": "", "result": "fresh", "reason": "explicit_fresh", "local_resume_path": null, "remote_repo": null, "remote_prefix": null, "compatibility": "unchecked", "config_fingerprint": "dda604a0a8933410f349c687646bc21b05e3367caf370e8732dcd7e8658a8569"}
2026-04-12 10:46:11,586 | 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 04-12 10:46:12 [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(

  0%|          | 1/3300 [00:12<11:23:16, 12.43s/it]2026-04-12 10:46:24,518 | INFO | train_grpo_train | metrics_logged mode=train step=1

                                                   
{'loss': 0.1041, 'grad_norm': 10.209015846252441, 'learning_rate': 1e-05, 'num_tokens': 2463.0, 'completions/mean_length': 121.875, 'completions/min_length': 81.0, 'completions/max_length': 160.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 121.875, 'completions/min_terminated_length': 81.0, 'completions/max_terminated_length': 160.0, 'rewards/meter/mean': 0.42592284083366394, 'rewards/meter/std': 0.33591756224632263, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.9937220811843872, 'rewards/repeat_penalty/std': 0.004517034627497196, 'rewards/near_duplicate_penalty/mean': 0.9937220811843872, 'rewards/near_duplicate_penalty/std': 0.004517034627497196, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.15128016471862793, 'rewards/total_composite/std': 0.13516460359096527, 'reward': 0.15128016471862793, 'reward_std': 0.13516458868980408, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2232884168624878, 'sampling/sampling_logp_difference/max': 2.2234439849853516, 'sampling/importance_sampling_ratio/min': 0.10823570936918259, 'sampling/importance_sampling_ratio/mean': 1.0212712287902832, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.127237632870674, 'clip_ratio/low_mean': 0.1246343944221735, 'clip_ratio/low_min': 0.1246343944221735, 'clip_ratio/high_mean': 0.08433323167264462, 'clip_ratio/high_max': 0.08433323167264462, 'clip_ratio/region_mean': 0.20896762609481812, 'reward_total_mean': 0.15128016471862793, 'reward_meter_mean': 0.42592284083366394, 'reward_meter_std': 0.33591756224632263, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9937220811843872, 'reward_repeat_penalty_std': 0.004517034627497196, 'reward_near_duplicate_penalty_mean': 0.9937220811843872, 'reward_near_duplicate_penalty_std': 0.004517034627497196, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.15128016471862793, 'reward_total_composite_std': 0.13516460359096527, 'epoch': 0.0}

  0%|          | 1/3300 [00:12<11:23:16, 12.43s/it]INFO 04-12 10:46:24 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 2/3300 [00:22<10:20:56, 11.30s/it]2026-04-12 10:46:34,984 | INFO | train_grpo_train | metrics_logged mode=train step=2

                                                   
{'loss': 0.105, 'grad_norm': 15.227958679199219, 'learning_rate': 9.996969696969698e-06, 'num_tokens': 4944.0, 'completions/mean_length': 109.125, 'completions/min_length': 96.0, 'completions/max_length': 129.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 109.125, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 129.0, 'rewards/meter/mean': 0.7208319902420044, 'rewards/meter/std': 0.2011304646730423, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7437499761581421, 'rewards/judge_quality/std': 0.2432481348514557, 'rewards/repeat_penalty/mean': 0.9835867881774902, 'rewards/repeat_penalty/std': 0.002223996678367257, 'rewards/near_duplicate_penalty/mean': 0.9835867881774902, 'rewards/near_duplicate_penalty/std': 0.002223996678367257, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.566516637802124, 'rewards/total_composite/std': 0.2884424030780792, 'reward': 0.566516637802124, 'reward_std': 0.2884424030780792, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15910686552524567, 'sampling/sampling_logp_difference/max': 3.482398509979248, 'sampling/importance_sampling_ratio/min': 0.030733609572052956, 'sampling/importance_sampling_ratio/mean': 0.9993525147438049, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.633471455425024, 'clip_ratio/low_mean': 0.06075856648385525, 'clip_ratio/low_min': 0.06075856648385525, 'clip_ratio/high_mean': 0.05482602585107088, 'clip_ratio/high_max': 0.05482602585107088, 'clip_ratio/region_mean': 0.11558459233492613, 'reward_total_mean': 0.566516637802124, 'reward_meter_mean': 0.7208319902420044, 'reward_meter_std': 0.2011304646730423, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9835867881774902, 'reward_repeat_penalty_std': 0.002223996678367257, 'reward_near_duplicate_penalty_mean': 0.9835867881774902, 'reward_near_duplicate_penalty_std': 0.002223996678367257, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7437499761581421, 'reward_judge_quality_std': 0.2432481348514557, 'reward_total_composite_mean': 0.566516637802124, 'reward_total_composite_std': 0.2884424030780792, 'epoch': 0.0}

  0%|          | 2/3300 [00:22<10:20:56, 11.30s/it]INFO 04-12 10:46:35 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 3/3300 [00:31<9:23:37, 10.26s/it] 2026-04-12 10:46:44,007 | INFO | train_grpo_train | metrics_logged mode=train step=3

                                                  
{'loss': 0.1276, 'grad_norm': 15.129196166992188, 'learning_rate': 9.993939393939395e-06, 'num_tokens': 6507.0, 'completions/mean_length': 40.375, 'completions/min_length': 23.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.375, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.5452406406402588, 'rewards/meter/std': 0.43102967739105225, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5299999713897705, 'rewards/judge_quality/std': 0.35063210129737854, 'rewards/repeat_penalty/mean': 0.99625563621521, 'rewards/repeat_penalty/std': 0.00466901296749711, 'rewards/near_duplicate_penalty/mean': 0.99625563621521, 'rewards/near_duplicate_penalty/std': 0.00466901296749711, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.24037756025791168, 'rewards/total_composite/std': 0.3203948438167572, 'reward': 0.24037756025791168, 'reward_std': 0.3203948438167572, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23142485320568085, 'sampling/sampling_logp_difference/max': 1.662959098815918, 'sampling/importance_sampling_ratio/min': 0.18957717716693878, 'sampling/importance_sampling_ratio/mean': 1.0225555896759033, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7431422844529152, 'clip_ratio/low_mean': 0.16276154574006796, 'clip_ratio/low_min': 0.16276154574006796, 'clip_ratio/high_mean': 0.05761563032865524, 'clip_ratio/high_max': 0.05761563032865524, 'clip_ratio/region_mean': 0.2203771760687232, 'reward_total_mean': 0.24037756025791168, 'reward_meter_mean': 0.5452406406402588, 'reward_meter_std': 0.43102967739105225, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.99625563621521, 'reward_repeat_penalty_std': 0.00466901296749711, 'reward_near_duplicate_penalty_mean': 0.99625563621521, 'reward_near_duplicate_penalty_std': 0.00466901296749711, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5299999713897705, 'reward_judge_quality_std': 0.35063210129737854, 'reward_total_composite_mean': 0.24037756025791168, 'reward_total_composite_std': 0.3203948438167572, 'epoch': 0.0}

  0%|          | 3/3300 [00:31<9:23:37, 10.26s/it]INFO 04-12 10:46:44 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 4/3300 [00:44<10:07:57, 11.07s/it]2026-04-12 10:46:56,295 | INFO | train_grpo_train | metrics_logged mode=train step=4

                                                   
{'loss': -0.0339, 'grad_norm': 8.383411407470703, 'learning_rate': 9.990909090909093e-06, 'num_tokens': 9567.0, 'completions/mean_length': 167.5, 'completions/min_length': 106.0, 'completions/max_length': 243.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 167.5, 'completions/min_terminated_length': 106.0, 'completions/max_terminated_length': 243.0, 'rewards/meter/mean': 0.3337455689907074, 'rewards/meter/std': 0.27989256381988525, 'rewards/count_adherence/mean': 0.9791666269302368, 'rewards/count_adherence/std': 0.0589255727827549, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9940738081932068, 'rewards/lexical_plausibility/std': 0.011887531727552414, 'rewards/judge_quality/mean': 0.3375000059604645, 'rewards/judge_quality/std': 0.08345229178667068, 'rewards/repeat_penalty/mean': 0.9878745675086975, 'rewards/repeat_penalty/std': 0.018476979807019234, 'rewards/near_duplicate_penalty/mean': 0.9936480522155762, 'rewards/near_duplicate_penalty/std': 0.006021037232130766, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9941724538803101, 'rewards/distinct_2/std': 0.016482684761285782, 'rewards/total_composite/mean': 0.09443728625774384, 'rewards/total_composite/std': 0.052527233958244324, 'reward': 0.09443728625774384, 'reward_std': 0.052527233958244324, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22497865557670593, 'sampling/sampling_logp_difference/max': 1.6787109375, 'sampling/importance_sampling_ratio/min': 0.18661437928676605, 'sampling/importance_sampling_ratio/mean': 1.0238019227981567, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.535755380988121, 'clip_ratio/low_mean': 0.04465997405350208, 'clip_ratio/low_min': 0.04465997405350208, 'clip_ratio/high_mean': 0.1727588288486004, 'clip_ratio/high_max': 0.1727588288486004, 'clip_ratio/region_mean': 0.21741880290210247, 'reward_total_mean': 0.09443728625774384, 'reward_meter_mean': 0.3337455689907074, 'reward_meter_std': 0.27989256381988525, 'reward_count_adherence_mean': 0.9791666269302368, 'reward_count_adherence_std': 0.0589255727827549, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9940738081932068, 'reward_lexical_plausibility_std': 0.011887531727552414, 'reward_repeat_penalty_mean': 0.9878745675086975, 'reward_repeat_penalty_std': 0.018476979807019234, 'reward_near_duplicate_penalty_mean': 0.9936480522155762, 'reward_near_duplicate_penalty_std': 0.006021037232130766, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9941724538803101, 'reward_distinct_2_std': 0.016482684761285782, 'reward_judge_quality_mean': 0.3375000059604645, 'reward_judge_quality_std': 0.08345229178667068, 'reward_total_composite_mean': 0.09443728625774384, 'reward_total_composite_std': 0.052527233958244324, 'epoch': 0.0}

  0%|          | 4/3300 [00:44<10:07:57, 11.07s/it]INFO 04-12 10:46:56 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 5/3300 [00:53<9:38:46, 10.54s/it] 2026-04-12 10:47:05,897 | INFO | train_grpo_train | metrics_logged mode=train step=5

                                                  
{'loss': 0.0498, 'grad_norm': 20.20990753173828, 'learning_rate': 9.987878787878788e-06, 'num_tokens': 11123.0, 'completions/mean_length': 40.5, 'completions/min_length': 31.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.5, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.29064032435417175, 'rewards/meter/std': 0.3684650957584381, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.47874999046325684, 'rewards/judge_quality/std': 0.1683056354522705, 'rewards/repeat_penalty/mean': 0.9905592799186707, 'rewards/repeat_penalty/std': 0.014576256275177002, 'rewards/near_duplicate_penalty/mean': 0.9905592799186707, 'rewards/near_duplicate_penalty/std': 0.014576256275177002, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.13773858547210693, 'rewards/total_composite/std': 0.16502845287322998, 'reward': 0.13773858547210693, 'reward_std': 0.1650284379720688, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2425910085439682, 'sampling/sampling_logp_difference/max': 1.4678411483764648, 'sampling/importance_sampling_ratio/min': 0.23042239248752594, 'sampling/importance_sampling_ratio/mean': 1.0286279916763306, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.3794576972723007, 'clip_ratio/low_mean': 0.17255527526140213, 'clip_ratio/low_min': 0.17255527526140213, 'clip_ratio/high_mean': 0.056785713881254196, 'clip_ratio/high_max': 0.056785713881254196, 'clip_ratio/region_mean': 0.22934098914265633, 'reward_total_mean': 0.13773858547210693, 'reward_meter_mean': 0.29064032435417175, 'reward_meter_std': 0.3684650957584381, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9905592799186707, 'reward_repeat_penalty_std': 0.014576256275177002, 'reward_near_duplicate_penalty_mean': 0.9905592799186707, 'reward_near_duplicate_penalty_std': 0.014576256275177002, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.47874999046325684, 'reward_judge_quality_std': 0.1683056354522705, 'reward_total_composite_mean': 0.13773858547210693, 'reward_total_composite_std': 0.16502845287322998, 'epoch': 0.0}

  0%|          | 5/3300 [00:53<9:38:46, 10.54s/it]INFO 04-12 10:47:06 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 6/3300 [01:02<9:11:09, 10.04s/it]2026-04-12 10:47:14,965 | INFO | train_grpo_train | metrics_logged mode=train step=6

                                                  
{'loss': -0.0723, 'grad_norm': 15.78359603881836, 'learning_rate': 9.984848484848485e-06, 'num_tokens': 12786.0, 'completions/mean_length': 49.875, 'completions/min_length': 34.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.875, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.5357946157455444, 'rewards/meter/std': 0.3965552747249603, '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/lexical_plausibility/mean': 0.990384578704834, 'rewards/lexical_plausibility/std': 0.01780424639582634, 'rewards/judge_quality/mean': 0.38374999165534973, 'rewards/judge_quality/std': 0.24076886475086212, 'rewards/repeat_penalty/mean': 0.9968063831329346, 'rewards/repeat_penalty/std': 0.005851962603628635, 'rewards/near_duplicate_penalty/mean': 0.9968063831329346, 'rewards/near_duplicate_penalty/std': 0.005851962603628635, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.19787782430648804, 'rewards/total_composite/std': 0.16528819501399994, 'reward': 0.19787782430648804, 'reward_std': 0.16528819501399994, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24047894775867462, 'sampling/sampling_logp_difference/max': 2.8058953285217285, 'sampling/importance_sampling_ratio/min': 0.060452621430158615, 'sampling/importance_sampling_ratio/mean': 1.0489442348480225, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.553137093782425, 'clip_ratio/low_mean': 0.11882871203124523, 'clip_ratio/low_min': 0.11882871203124523, 'clip_ratio/high_mean': 0.10576356761157513, 'clip_ratio/high_max': 0.10576356761157513, 'clip_ratio/region_mean': 0.22459227964282036, 'reward_total_mean': 0.19787782430648804, 'reward_meter_mean': 0.5357946157455444, 'reward_meter_std': 0.3965552747249603, '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_lexical_plausibility_mean': 0.990384578704834, 'reward_lexical_plausibility_std': 0.01780424639582634, 'reward_repeat_penalty_mean': 0.9968063831329346, 'reward_repeat_penalty_std': 0.005851962603628635, 'reward_near_duplicate_penalty_mean': 0.9968063831329346, 'reward_near_duplicate_penalty_std': 0.005851962603628635, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.38374999165534973, 'reward_judge_quality_std': 0.24076886475086212, 'reward_total_composite_mean': 0.19787782430648804, 'reward_total_composite_std': 0.16528819501399994, 'epoch': 0.0}

  0%|          | 6/3300 [01:02<9:11:09, 10.04s/it]INFO 04-12 10:47:15 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 7/3300 [01:11<8:46:17,  9.59s/it]2026-04-12 10:47:23,629 | INFO | train_grpo_train | metrics_logged mode=train step=7

                                                  
{'loss': 0.1518, 'grad_norm': 19.81330108642578, 'learning_rate': 9.981818181818183e-06, 'num_tokens': 14380.0, 'completions/mean_length': 40.25, 'completions/min_length': 34.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.25, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.561120867729187, 'rewards/meter/std': 0.38149482011795044, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5337499976158142, 'rewards/judge_quality/std': 0.21540245413780212, 'rewards/repeat_penalty/mean': 0.9547847509384155, 'rewards/repeat_penalty/std': 0.1135256439447403, 'rewards/near_duplicate_penalty/mean': 0.9917590618133545, 'rewards/near_duplicate_penalty/std': 0.01057454850524664, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9908424615859985, 'rewards/distinct_2/std': 0.025901345536112785, 'rewards/total_composite/mean': 0.28039056062698364, 'rewards/total_composite/std': 0.21561163663864136, 'reward': 0.28039056062698364, 'reward_std': 0.21561162173748016, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21762706339359283, 'sampling/sampling_logp_difference/max': 1.4331035614013672, 'sampling/importance_sampling_ratio/min': 0.23856738209724426, 'sampling/importance_sampling_ratio/mean': 1.0254625082015991, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9640412628650665, 'clip_ratio/low_mean': 0.08554505370557308, 'clip_ratio/low_min': 0.08554505370557308, 'clip_ratio/high_mean': 0.10183645226061344, 'clip_ratio/high_max': 0.10183645226061344, 'clip_ratio/region_mean': 0.18738150596618652, 'reward_total_mean': 0.28039056062698364, 'reward_meter_mean': 0.561120867729187, 'reward_meter_std': 0.38149482011795044, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9547847509384155, 'reward_repeat_penalty_std': 0.1135256439447403, 'reward_near_duplicate_penalty_mean': 0.9917590618133545, 'reward_near_duplicate_penalty_std': 0.01057454850524664, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9908424615859985, 'reward_distinct_2_std': 0.025901345536112785, 'reward_judge_quality_mean': 0.5337499976158142, 'reward_judge_quality_std': 0.21540245413780212, 'reward_total_composite_mean': 0.28039056062698364, 'reward_total_composite_std': 0.21561163663864136, 'epoch': 0.0}

  0%|          | 7/3300 [01:11<8:46:17,  9.59s/it]INFO 04-12 10:47:23 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 8/3300 [01:25<9:54:37, 10.84s/it]2026-04-12 10:47:37,140 | INFO | train_grpo_train | metrics_logged mode=train step=8

                                                  
{'loss': -0.0938, 'grad_norm': 6.824868679046631, 'learning_rate': 9.97878787878788e-06, 'num_tokens': 17989.0, 'completions/mean_length': 256.125, 'completions/min_length': 99.0, 'completions/max_length': 396.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 256.125, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 396.0, 'rewards/meter/mean': 0.6256775856018066, 'rewards/meter/std': 0.17330610752105713, 'rewards/count_adherence/mean': 0.8977273106575012, 'rewards/count_adherence/std': 0.14928004145622253, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.9633318185806274, 'rewards/repeat_penalty/std': 0.058776505291461945, 'rewards/near_duplicate_penalty/mean': 0.9902274012565613, 'rewards/near_duplicate_penalty/std': 0.012147120200097561, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9723308086395264, 'rewards/distinct_2/std': 0.04843715578317642, 'rewards/total_composite/mean': 0.17009173333644867, 'rewards/total_composite/std': 0.10656578093767166, 'reward': 0.17009173333644867, 'reward_std': 0.10656578093767166, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22988878190517426, 'sampling/sampling_logp_difference/max': 1.8197107315063477, 'sampling/importance_sampling_ratio/min': 0.16207262873649597, 'sampling/importance_sampling_ratio/mean': 1.0383141040802002, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.8317363262176514, 'clip_ratio/low_mean': 0.07635685615241528, 'clip_ratio/low_min': 0.07635685615241528, 'clip_ratio/high_mean': 0.1352237407118082, 'clip_ratio/high_max': 0.1352237407118082, 'clip_ratio/region_mean': 0.21158059686422348, 'reward_total_mean': 0.17009173333644867, 'reward_meter_mean': 0.6256775856018066, 'reward_meter_std': 0.17330610752105713, 'reward_count_adherence_mean': 0.8977273106575012, 'reward_count_adherence_std': 0.14928004145622253, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9633318185806274, 'reward_repeat_penalty_std': 0.058776505291461945, 'reward_near_duplicate_penalty_mean': 0.9902274012565613, 'reward_near_duplicate_penalty_std': 0.012147120200097561, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9723308086395264, 'reward_distinct_2_std': 0.04843715578317642, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.17009173333644867, 'reward_total_composite_std': 0.10656578093767166, 'epoch': 0.0}

  0%|          | 8/3300 [01:25<9:54:37, 10.84s/it]INFO 04-12 10:47:37 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 9/3300 [01:35<9:51:56, 10.79s/it]2026-04-12 10:47:47,832 | INFO | train_grpo_train | metrics_logged mode=train step=9

                                                  
{'loss': 0.0476, 'grad_norm': 18.376598358154297, 'learning_rate': 9.975757575757577e-06, 'num_tokens': 20445.0, 'completions/mean_length': 124.0, 'completions/min_length': 85.0, 'completions/max_length': 146.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 124.0, 'completions/min_terminated_length': 85.0, 'completions/max_terminated_length': 146.0, 'rewards/meter/mean': 0.7044054269790649, 'rewards/meter/std': 0.32541099190711975, 'rewards/count_adherence/mean': 0.9583333134651184, 'rewards/count_adherence/std': 0.07715168595314026, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6474999785423279, 'rewards/judge_quality/std': 0.297981321811676, 'rewards/repeat_penalty/mean': 0.9833334684371948, 'rewards/repeat_penalty/std': 0.0076095727272331715, 'rewards/near_duplicate_penalty/mean': 0.9833334684371948, 'rewards/near_duplicate_penalty/std': 0.0076095727272331715, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.49316665530204773, 'rewards/total_composite/std': 0.3613712787628174, 'reward': 0.49316665530204773, 'reward_std': 0.36137130856513977, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19040951132774353, 'sampling/sampling_logp_difference/max': 2.7353835105895996, 'sampling/importance_sampling_ratio/min': 0.06486912816762924, 'sampling/importance_sampling_ratio/mean': 0.9912941455841064, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7231591157615185, 'clip_ratio/low_mean': 0.1145949698984623, 'clip_ratio/low_min': 0.1145949698984623, 'clip_ratio/high_mean': 0.0339728519320488, 'clip_ratio/high_max': 0.0339728519320488, 'clip_ratio/region_mean': 0.1485678218305111, 'reward_total_mean': 0.49316665530204773, 'reward_meter_mean': 0.7044054269790649, 'reward_meter_std': 0.32541099190711975, 'reward_count_adherence_mean': 0.9583333134651184, 'reward_count_adherence_std': 0.07715168595314026, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9833334684371948, 'reward_repeat_penalty_std': 0.0076095727272331715, 'reward_near_duplicate_penalty_mean': 0.9833334684371948, 'reward_near_duplicate_penalty_std': 0.0076095727272331715, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6474999785423279, 'reward_judge_quality_std': 0.297981321811676, 'reward_total_composite_mean': 0.49316665530204773, 'reward_total_composite_std': 0.3613712787628174, 'epoch': 0.0}

  0%|          | 9/3300 [01:35<9:51:56, 10.79s/it]INFO 04-12 10:47:48 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 10/3300 [01:46<9:50:33, 10.77s/it]2026-04-12 10:47:58,553 | INFO | train_grpo_train | metrics_logged mode=train step=10

                                                   
{'loss': 0.2127, 'grad_norm': 12.192911148071289, 'learning_rate': 9.972727272727274e-06, 'num_tokens': 22691.0, 'completions/mean_length': 106.75, 'completions/min_length': 72.0, 'completions/max_length': 139.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 106.75, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 139.0, 'rewards/meter/mean': 0.6715837717056274, 'rewards/meter/std': 0.3977217972278595, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.10690449178218842, 'rewards/repeat_penalty/mean': 0.9872806668281555, 'rewards/repeat_penalty/std': 0.013390115462243557, 'rewards/near_duplicate_penalty/mean': 0.9872806668281555, 'rewards/near_duplicate_penalty/std': 0.013390115462243557, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.25240373611450195, 'rewards/total_composite/std': 0.16661977767944336, 'reward': 0.25240373611450195, 'reward_std': 0.16661976277828217, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20312635600566864, 'sampling/sampling_logp_difference/max': 1.5351839065551758, 'sampling/importance_sampling_ratio/min': 0.2154160737991333, 'sampling/importance_sampling_ratio/mean': 1.0280596017837524, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.019076704978943, 'clip_ratio/low_mean': 0.09650039486587048, 'clip_ratio/low_min': 0.09650039486587048, 'clip_ratio/high_mean': 0.10350521840155125, 'clip_ratio/high_max': 0.10350521840155125, 'clip_ratio/region_mean': 0.20000561326742172, 'reward_total_mean': 0.25240373611450195, 'reward_meter_mean': 0.6715837717056274, 'reward_meter_std': 0.3977217972278595, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9872806668281555, 'reward_repeat_penalty_std': 0.013390115462243557, 'reward_near_duplicate_penalty_mean': 0.9872806668281555, 'reward_near_duplicate_penalty_std': 0.013390115462243557, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.10690449178218842, 'reward_total_composite_mean': 0.25240373611450195, 'reward_total_composite_std': 0.16661977767944336, 'epoch': 0.0}

  0%|          | 10/3300 [01:46<9:50:33, 10.77s/it]INFO 04-12 10:47:58 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 11/3300 [01:55<9:22:22, 10.26s/it]2026-04-12 10:48:07,654 | INFO | train_grpo_train | metrics_logged mode=train step=11

                                                   
{'loss': 0.0765, 'grad_norm': 13.076592445373535, 'learning_rate': 9.96969696969697e-06, 'num_tokens': 24463.0, 'completions/mean_length': 55.5, 'completions/min_length': 46.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.5, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.27222830057144165, 'rewards/meter/std': 0.2973862588405609, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9712885022163391, 'rewards/lexical_plausibility/std': 0.08120834082365036, 'rewards/judge_quality/mean': 0.42124998569488525, 'rewards/judge_quality/std': 0.22177450358867645, 'rewards/repeat_penalty/mean': 0.9347968101501465, 'rewards/repeat_penalty/std': 0.17061248421669006, 'rewards/near_duplicate_penalty/mean': 0.9826840162277222, 'rewards/near_duplicate_penalty/std': 0.03586622327566147, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9704141616821289, 'rewards/distinct_2/std': 0.08368128538131714, 'rewards/total_composite/mean': 0.16706141829490662, 'rewards/total_composite/std': 0.2936072051525116, 'reward': 0.16706141829490662, 'reward_std': 0.2936071753501892, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2019934505224228, 'sampling/sampling_logp_difference/max': 2.018817901611328, 'sampling/importance_sampling_ratio/min': 0.13281236588954926, 'sampling/importance_sampling_ratio/mean': 1.0267703533172607, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3748232424259186, 'clip_ratio/low_mean': 0.14291187189519405, 'clip_ratio/low_min': 0.14291187189519405, 'clip_ratio/high_mean': 0.05877896584570408, 'clip_ratio/high_max': 0.05877896584570408, 'clip_ratio/region_mean': 0.20169083774089813, 'reward_total_mean': 0.16706141829490662, 'reward_meter_mean': 0.27222830057144165, 'reward_meter_std': 0.2973862588405609, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9712885022163391, 'reward_lexical_plausibility_std': 0.08120834082365036, 'reward_repeat_penalty_mean': 0.9347968101501465, 'reward_repeat_penalty_std': 0.17061248421669006, 'reward_near_duplicate_penalty_mean': 0.9826840162277222, 'reward_near_duplicate_penalty_std': 0.03586622327566147, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9704141616821289, 'reward_distinct_2_std': 0.08368128538131714, 'reward_judge_quality_mean': 0.42124998569488525, 'reward_judge_quality_std': 0.22177450358867645, 'reward_total_composite_mean': 0.16706141829490662, 'reward_total_composite_std': 0.2936072051525116, 'epoch': 0.0}

  0%|          | 11/3300 [01:55<9:22:22, 10.26s/it]INFO 04-12 10:48:07 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 12/3300 [02:10<10:34:52, 11.59s/it]2026-04-12 10:48:22,272 | INFO | train_grpo_train | metrics_logged mode=train step=12

                                                    
{'loss': 0.0813, 'grad_norm': 5.912014961242676, 'learning_rate': 9.966666666666667e-06, 'num_tokens': 29097.0, 'completions/mean_length': 354.25, 'completions/min_length': 201.0, 'completions/max_length': 487.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 354.25, 'completions/min_terminated_length': 201.0, 'completions/max_terminated_length': 487.0, 'rewards/meter/mean': 0.35211530327796936, 'rewards/meter/std': 0.24533754587173462, 'rewards/count_adherence/mean': 0.884615421295166, 'rewards/count_adherence/std': 0.058148376643657684, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.23749999701976776, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.9808359146118164, 'rewards/repeat_penalty/std': 0.016977891325950623, 'rewards/near_duplicate_penalty/mean': 0.9925937652587891, 'rewards/near_duplicate_penalty/std': 0.004112385679036379, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9881381988525391, 'rewards/distinct_2/std': 0.015464049763977528, 'rewards/total_composite/mean': 0.0739469900727272, 'rewards/total_composite/std': 0.054342567920684814, 'reward': 0.0739469900727272, 'reward_std': 0.054342564195394516, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22679691016674042, 'sampling/sampling_logp_difference/max': 2.1510400772094727, 'sampling/importance_sampling_ratio/min': 0.11636307090520859, 'sampling/importance_sampling_ratio/mean': 1.038204550743103, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.89529550075531, 'clip_ratio/low_mean': 0.11654424481093884, 'clip_ratio/low_min': 0.11654424481093884, 'clip_ratio/high_mean': 0.08761620707809925, 'clip_ratio/high_max': 0.08761620707809925, 'clip_ratio/region_mean': 0.20416045188903809, 'reward_total_mean': 0.0739469900727272, 'reward_meter_mean': 0.35211530327796936, 'reward_meter_std': 0.24533754587173462, 'reward_count_adherence_mean': 0.884615421295166, 'reward_count_adherence_std': 0.058148376643657684, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9808359146118164, 'reward_repeat_penalty_std': 0.016977891325950623, 'reward_near_duplicate_penalty_mean': 0.9925937652587891, 'reward_near_duplicate_penalty_std': 0.004112385679036379, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9881381988525391, 'reward_distinct_2_std': 0.015464049763977528, 'reward_judge_quality_mean': 0.23749999701976776, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.0739469900727272, 'reward_total_composite_std': 0.054342567920684814, 'epoch': 0.0}

  0%|          | 12/3300 [02:10<10:34:52, 11.59s/it]INFO 04-12 10:48:22 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 13/3300 [02:20<10:10:48, 11.15s/it]2026-04-12 10:48:32,419 | INFO | train_grpo_train | metrics_logged mode=train step=13

                                                    
{'loss': 0.2481, 'grad_norm': 19.421398162841797, 'learning_rate': 9.963636363636364e-06, 'num_tokens': 30718.0, 'completions/mean_length': 42.625, 'completions/min_length': 31.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.625, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.5187985897064209, 'rewards/meter/std': 0.4365939795970917, '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/lexical_plausibility/mean': 0.8742079734802246, 'rewards/lexical_plausibility/std': 0.19562670588493347, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.14078859984874725, 'rewards/repeat_penalty/mean': 0.9942833185195923, 'rewards/repeat_penalty/std': 0.01435972098261118, 'rewards/near_duplicate_penalty/mean': 0.9942833185195923, 'rewards/near_duplicate_penalty/std': 0.01435972098261118, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.15214984118938446, 'rewards/total_composite/std': 0.18843550980091095, 'reward': 0.15214984118938446, 'reward_std': 0.18843549489974976, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23536482453346252, 'sampling/sampling_logp_difference/max': 2.879500389099121, 'sampling/importance_sampling_ratio/min': 0.056162815541028976, 'sampling/importance_sampling_ratio/mean': 1.0110152959823608, 'sampling/importance_sampling_ratio/max': 1.987526774406433, 'entropy': 1.822215810418129, 'clip_ratio/low_mean': 0.10695216665044427, 'clip_ratio/low_min': 0.10695216665044427, 'clip_ratio/high_mean': 0.08850806392729282, 'clip_ratio/high_max': 0.08850806392729282, 'clip_ratio/region_mean': 0.1954602305777371, 'reward_total_mean': 0.15214984118938446, 'reward_meter_mean': 0.5187985897064209, 'reward_meter_std': 0.4365939795970917, '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_lexical_plausibility_mean': 0.8742079734802246, 'reward_lexical_plausibility_std': 0.19562670588493347, 'reward_repeat_penalty_mean': 0.9942833185195923, 'reward_repeat_penalty_std': 0.01435972098261118, 'reward_near_duplicate_penalty_mean': 0.9942833185195923, 'reward_near_duplicate_penalty_std': 0.01435972098261118, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.14078859984874725, 'reward_total_composite_mean': 0.15214984118938446, 'reward_total_composite_std': 0.18843550980091095, 'epoch': 0.0}

  0%|          | 13/3300 [02:20<10:10:48, 11.15s/it]INFO 04-12 10:48:32 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 14/3300 [02:28<9:27:18, 10.36s/it] 2026-04-12 10:48:40,950 | INFO | train_grpo_train | metrics_logged mode=train step=14

                                                   
{'loss': 0.0002, 'grad_norm': 16.93572998046875, 'learning_rate': 9.960606060606062e-06, 'num_tokens': 32314.0, 'completions/mean_length': 40.5, 'completions/min_length': 31.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.5, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.8533340692520142, 'rewards/meter/std': 0.20150882005691528, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.44999998807907104, 'rewards/judge_quality/std': 0.09258200973272324, 'rewards/repeat_penalty/mean': 0.9495680928230286, 'rewards/repeat_penalty/std': 0.11365896463394165, 'rewards/near_duplicate_penalty/mean': 0.9868184328079224, 'rewards/near_duplicate_penalty/std': 0.016084030270576477, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.990384578704834, 'rewards/distinct_2/std': 0.027196412906050682, 'rewards/total_composite/mean': 0.37042945623397827, 'rewards/total_composite/std': 0.06580816209316254, 'reward': 0.37042945623397827, 'reward_std': 0.06580816209316254, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19301632046699524, 'sampling/sampling_logp_difference/max': 1.195810317993164, 'sampling/importance_sampling_ratio/min': 0.3146522641181946, 'sampling/importance_sampling_ratio/mean': 1.0376253128051758, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7731505632400513, 'clip_ratio/low_mean': 0.09220211394131184, 'clip_ratio/low_min': 0.09220211394131184, 'clip_ratio/high_mean': 0.08251201920211315, 'clip_ratio/high_max': 0.08251201920211315, 'clip_ratio/region_mean': 0.174714133143425, 'reward_total_mean': 0.37042945623397827, 'reward_meter_mean': 0.8533340692520142, 'reward_meter_std': 0.20150882005691528, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.9495680928230286, 'reward_repeat_penalty_std': 0.11365896463394165, 'reward_near_duplicate_penalty_mean': 0.9868184328079224, 'reward_near_duplicate_penalty_std': 0.016084030270576477, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.990384578704834, 'reward_distinct_2_std': 0.027196412906050682, 'reward_judge_quality_mean': 0.44999998807907104, 'reward_judge_quality_std': 0.09258200973272324, 'reward_total_composite_mean': 0.37042945623397827, 'reward_total_composite_std': 0.06580816209316254, 'epoch': 0.0}

  0%|          | 14/3300 [02:28<9:27:18, 10.36s/it]INFO 04-12 10:48:41 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 15/3300 [02:38<9:13:10, 10.10s/it]2026-04-12 10:48:50,462 | INFO | train_grpo_train | metrics_logged mode=train step=15

                                                   
{'loss': 0.0265, 'grad_norm': 16.918575286865234, 'learning_rate': 9.957575757575757e-06, 'num_tokens': 34037.0, 'completions/mean_length': 45.375, 'completions/min_length': 38.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.375, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.6372770071029663, 'rewards/meter/std': 0.43653595447540283, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9914492964744568, 'rewards/lexical_plausibility/std': 0.024185042828321457, 'rewards/judge_quality/mean': 0.5174999833106995, 'rewards/judge_quality/std': 0.2522329092025757, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.30071666836738586, 'rewards/total_composite/std': 0.27117469906806946, 'reward': 0.30071666836738586, 'reward_std': 0.27117466926574707, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.224419966340065, 'sampling/sampling_logp_difference/max': 1.8789808750152588, 'sampling/importance_sampling_ratio/min': 0.17696641385555267, 'sampling/importance_sampling_ratio/mean': 1.0451421737670898, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2488113194704056, 'clip_ratio/low_mean': 0.07137374579906464, 'clip_ratio/low_min': 0.07137374579906464, 'clip_ratio/high_mean': 0.14608749747276306, 'clip_ratio/high_max': 0.14608749747276306, 'clip_ratio/region_mean': 0.2174612432718277, 'reward_total_mean': 0.30071666836738586, 'reward_meter_mean': 0.6372770071029663, 'reward_meter_std': 0.43653595447540283, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9914492964744568, 'reward_lexical_plausibility_std': 0.024185042828321457, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5174999833106995, 'reward_judge_quality_std': 0.2522329092025757, 'reward_total_composite_mean': 0.30071666836738586, 'reward_total_composite_std': 0.27117469906806946, 'epoch': 0.0}

  0%|          | 15/3300 [02:38<9:13:10, 10.10s/it]INFO 04-12 10:48:50 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 16/3300 [02:53<10:36:17, 11.63s/it]2026-04-12 10:49:05,621 | INFO | train_grpo_train | metrics_logged mode=train step=16

                                                    
{'loss': -0.1238, 'grad_norm': 4.117391109466553, 'learning_rate': 9.954545454545456e-06, 'num_tokens': 37943.0, 'completions/mean_length': 341.25, 'completions/min_length': 241.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 316.8571472167969, 'completions/min_terminated_length': 241.0, 'completions/max_terminated_length': 505.0, 'rewards/meter/mean': 0.7447863817214966, 'rewards/meter/std': 0.20929576456546783, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.0629940927028656, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.987236499786377, 'rewards/lexical_plausibility/std': 0.03610069304704666, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/repeat_penalty/mean': 0.9948620796203613, 'rewards/repeat_penalty/std': 0.0043237595818936825, 'rewards/near_duplicate_penalty/mean': 0.9962705373764038, 'rewards/near_duplicate_penalty/std': 0.001476338948123157, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9985859394073486, 'rewards/distinct_2/std': 0.003999483305960894, 'rewards/total_composite/mean': 0.21612676978111267, 'rewards/total_composite/std': 0.08605542778968811, 'reward': 0.21612676978111267, 'reward_std': 0.08605542778968811, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.226578027009964, 'sampling/sampling_logp_difference/max': 1.4890222549438477, 'sampling/importance_sampling_ratio/min': 0.22559311985969543, 'sampling/importance_sampling_ratio/mean': 1.0435864925384521, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.8357963263988495, 'clip_ratio/low_mean': 0.04257149621844292, 'clip_ratio/low_min': 0.04257149621844292, 'clip_ratio/high_mean': 0.1372001487761736, 'clip_ratio/high_max': 0.1372001487761736, 'clip_ratio/region_mean': 0.1797716449946165, 'reward_total_mean': 0.21612676978111267, 'reward_meter_mean': 0.7447863817214966, 'reward_meter_std': 0.20929576456546783, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.0629940927028656, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.987236499786377, 'reward_lexical_plausibility_std': 0.03610069304704666, 'reward_repeat_penalty_mean': 0.9948620796203613, 'reward_repeat_penalty_std': 0.0043237595818936825, 'reward_near_duplicate_penalty_mean': 0.9962705373764038, 'reward_near_duplicate_penalty_std': 0.001476338948123157, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9985859394073486, 'reward_distinct_2_std': 0.003999483305960894, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.21612676978111267, 'reward_total_composite_std': 0.08605542778968811, 'epoch': 0.0}

  0%|          | 16/3300 [02:53<10:36:17, 11.63s/it]INFO 04-12 10:49:05 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 17/3300 [03:02<9:48:35, 10.76s/it] 2026-04-12 10:49:14,359 | INFO | train_grpo_train | metrics_logged mode=train step=17

                                                   
{'loss': 0.1153, 'grad_norm': 12.76971435546875, 'learning_rate': 9.951515151515152e-06, 'num_tokens': 39853.0, 'completions/mean_length': 61.75, 'completions/min_length': 52.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.75, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.5454095602035522, 'rewards/meter/std': 0.4127260148525238, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9719981551170349, 'rewards/lexical_plausibility/std': 0.06866878271102905, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.15059404075145721, 'rewards/repeat_penalty/mean': 0.9964045882225037, 'rewards/repeat_penalty/std': 0.009191551245748997, 'rewards/near_duplicate_penalty/mean': 0.9964045882225037, 'rewards/near_duplicate_penalty/std': 0.009191551245748997, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.15721666812896729, 'rewards/total_composite/std': 0.17351087927818298, 'reward': 0.15721666812896729, 'reward_std': 0.17351087927818298, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21241459250450134, 'sampling/sampling_logp_difference/max': 1.8289339542388916, 'sampling/importance_sampling_ratio/min': 0.16058465838432312, 'sampling/importance_sampling_ratio/mean': 1.0272700786590576, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.273869574069977, 'clip_ratio/low_mean': 0.1418552715331316, 'clip_ratio/low_min': 0.1418552715331316, 'clip_ratio/high_mean': 0.05208333395421505, 'clip_ratio/high_max': 0.05208333395421505, 'clip_ratio/region_mean': 0.19393860548734665, 'reward_total_mean': 0.15721666812896729, 'reward_meter_mean': 0.5454095602035522, 'reward_meter_std': 0.4127260148525238, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9719981551170349, 'reward_lexical_plausibility_std': 0.06866878271102905, 'reward_repeat_penalty_mean': 0.9964045882225037, 'reward_repeat_penalty_std': 0.009191551245748997, 'reward_near_duplicate_penalty_mean': 0.9964045882225037, 'reward_near_duplicate_penalty_std': 0.009191551245748997, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.15059404075145721, 'reward_total_composite_mean': 0.15721666812896729, 'reward_total_composite_std': 0.17351087927818298, 'epoch': 0.0}

  1%|          | 17/3300 [03:02<9:48:35, 10.76s/it]INFO 04-12 10:49:14 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 18/3300 [03:12<9:36:32, 10.54s/it]2026-04-12 10:49:24,394 | INFO | train_grpo_train | metrics_logged mode=train step=18

                                                   
{'loss': 0.1304, 'grad_norm': 10.411269187927246, 'learning_rate': 9.948484848484849e-06, 'num_tokens': 42337.0, 'completions/mean_length': 111.5, 'completions/min_length': 82.0, 'completions/max_length': 153.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 111.5, 'completions/min_terminated_length': 82.0, 'completions/max_terminated_length': 153.0, 'rewards/meter/mean': 0.30559873580932617, 'rewards/meter/std': 0.2158714085817337, '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/lexical_plausibility/mean': 0.9970646500587463, 'rewards/lexical_plausibility/std': 0.008302465081214905, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.993833065032959, 'rewards/repeat_penalty/std': 0.005085999146103859, 'rewards/near_duplicate_penalty/mean': 0.993833065032959, 'rewards/near_duplicate_penalty/std': 0.005085999146103859, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.09969823062419891, 'rewards/total_composite/std': 0.07957087457180023, 'reward': 0.09969823062419891, 'reward_std': 0.07957086712121964, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22139480710029602, 'sampling/sampling_logp_difference/max': 1.6972432136535645, 'sampling/importance_sampling_ratio/min': 0.1831878274679184, 'sampling/importance_sampling_ratio/mean': 1.0457409620285034, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.325631469488144, 'clip_ratio/low_mean': 0.1352362111210823, 'clip_ratio/low_min': 0.1352362111210823, 'clip_ratio/high_mean': 0.07688837312161922, 'clip_ratio/high_max': 0.07688837312161922, 'clip_ratio/region_mean': 0.21212458424270153, 'reward_total_mean': 0.09969823062419891, 'reward_meter_mean': 0.30559873580932617, 'reward_meter_std': 0.2158714085817337, '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_lexical_plausibility_mean': 0.9970646500587463, 'reward_lexical_plausibility_std': 0.008302465081214905, 'reward_repeat_penalty_mean': 0.993833065032959, 'reward_repeat_penalty_std': 0.005085999146103859, 'reward_near_duplicate_penalty_mean': 0.993833065032959, 'reward_near_duplicate_penalty_std': 0.005085999146103859, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.09969823062419891, 'reward_total_composite_std': 0.07957087457180023, 'epoch': 0.0}

  1%|          | 18/3300 [03:12<9:36:32, 10.54s/it]INFO 04-12 10:49:24 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 19/3300 [03:27<10:56:01, 12.00s/it]2026-04-12 10:49:39,785 | INFO | train_grpo_train | metrics_logged mode=train step=19

                                                    
{'loss': -0.1513, 'grad_norm': 4.030255317687988, 'learning_rate': 9.945454545454546e-06, 'num_tokens': 45873.0, 'completions/mean_length': 423.0, 'completions/min_length': 340.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 369.6000061035156, 'completions/min_terminated_length': 340.0, 'completions/max_terminated_length': 421.0, 'rewards/meter/mean': 0.7693527936935425, 'rewards/meter/std': 0.2262124866247177, 'rewards/count_adherence/mean': 0.859375, 'rewards/count_adherence/std': 0.09300298243761063, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.23750001192092896, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.993139922618866, 'rewards/repeat_penalty/std': 0.004596802871674299, 'rewards/near_duplicate_penalty/mean': 0.9956949353218079, 'rewards/near_duplicate_penalty/std': 0.0011396065820008516, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9974359273910522, 'rewards/distinct_2/std': 0.004941675346344709, 'rewards/total_composite/mean': 0.12036975473165512, 'rewards/total_composite/std': 0.1208464652299881, 'reward': 0.12036975473165512, 'reward_std': 0.1208464577794075, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2371300756931305, 'sampling/sampling_logp_difference/max': 1.7566866874694824, 'sampling/importance_sampling_ratio/min': 0.17261584103107452, 'sampling/importance_sampling_ratio/mean': 1.0483399629592896, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2320169508457184, 'clip_ratio/low_mean': 0.044943422079086304, 'clip_ratio/low_min': 0.044943422079086304, 'clip_ratio/high_mean': 0.07177441194653511, 'clip_ratio/high_max': 0.07177441194653511, 'clip_ratio/region_mean': 0.11671783402562141, 'reward_total_mean': 0.12036975473165512, 'reward_meter_mean': 0.7693527936935425, 'reward_meter_std': 0.2262124866247177, 'reward_count_adherence_mean': 0.859375, 'reward_count_adherence_std': 0.09300298243761063, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.993139922618866, 'reward_repeat_penalty_std': 0.004596802871674299, 'reward_near_duplicate_penalty_mean': 0.9956949353218079, 'reward_near_duplicate_penalty_std': 0.0011396065820008516, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9974359273910522, 'reward_distinct_2_std': 0.004941675346344709, 'reward_judge_quality_mean': 0.23750001192092896, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.12036975473165512, 'reward_total_composite_std': 0.1208464652299881, 'epoch': 0.0}

  1%|          | 19/3300 [03:27<10:56:01, 12.00s/it]INFO 04-12 10:49:40 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 20/3300 [03:37<10:15:22, 11.26s/it]2026-04-12 10:49:49,317 | INFO | train_grpo_train | metrics_logged mode=train step=20

                                                    
{'loss': -0.0446, 'grad_norm': 14.589115142822266, 'learning_rate': 9.942424242424244e-06, 'num_tokens': 47559.0, 'completions/mean_length': 56.75, 'completions/min_length': 50.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.75, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.7581713199615479, 'rewards/meter/std': 0.27885109186172485, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4375, 'rewards/judge_quality/std': 0.1642080396413803, 'rewards/repeat_penalty/mean': 0.953586757183075, 'rewards/repeat_penalty/std': 0.05533900484442711, 'rewards/near_duplicate_penalty/mean': 0.984272837638855, 'rewards/near_duplicate_penalty/std': 0.017149362713098526, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9687758684158325, 'rewards/distinct_2/std': 0.052724964916706085, 'rewards/total_composite/mean': 0.313514769077301, 'rewards/total_composite/std': 0.17180955410003662, 'reward': 0.313514769077301, 'reward_std': 0.17180955410003662, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20675396919250488, 'sampling/sampling_logp_difference/max': 1.9559001922607422, 'sampling/importance_sampling_ratio/min': 0.1414370983839035, 'sampling/importance_sampling_ratio/mean': 1.0213472843170166, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1328125447034836, 'clip_ratio/low_mean': 0.07940024230629206, 'clip_ratio/low_min': 0.07940024230629206, 'clip_ratio/high_mean': 0.12832210678607225, 'clip_ratio/high_max': 0.12832210678607225, 'clip_ratio/region_mean': 0.2077223490923643, 'reward_total_mean': 0.313514769077301, 'reward_meter_mean': 0.7581713199615479, 'reward_meter_std': 0.27885109186172485, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.953586757183075, 'reward_repeat_penalty_std': 0.05533900484442711, 'reward_near_duplicate_penalty_mean': 0.984272837638855, 'reward_near_duplicate_penalty_std': 0.017149362713098526, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9687758684158325, 'reward_distinct_2_std': 0.052724964916706085, 'reward_judge_quality_mean': 0.4375, 'reward_judge_quality_std': 0.1642080396413803, 'reward_total_composite_mean': 0.313514769077301, 'reward_total_composite_std': 0.17180955410003662, 'epoch': 0.0}

  1%|          | 20/3300 [03:37<10:15:22, 11.26s/it]INFO 04-12 10:49:49 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 21/3300 [03:46<9:34:29, 10.51s/it] 2026-04-12 10:49:58,093 | INFO | train_grpo_train | metrics_logged mode=train step=21

                                                   
{'loss': -0.1181, 'grad_norm': 23.028400421142578, 'learning_rate': 9.939393939393939e-06, 'num_tokens': 49413.0, 'completions/mean_length': 63.75, 'completions/min_length': 46.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.75, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.30991411209106445, 'rewards/meter/std': 0.31503260135650635, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.42499998211860657, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.951915979385376, 'rewards/repeat_penalty/std': 0.04889478534460068, 'rewards/near_duplicate_penalty/mean': 0.9737776517868042, 'rewards/near_duplicate_penalty/std': 0.0162584837526083, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9773937463760376, 'rewards/distinct_2/std': 0.04370811954140663, 'rewards/total_composite/mean': 0.11938978731632233, 'rewards/total_composite/std': 0.13373103737831116, 'reward': 0.11938978731632233, 'reward_std': 0.13373102247714996, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.25628936290740967, 'sampling/sampling_logp_difference/max': 2.406100273132324, 'sampling/importance_sampling_ratio/min': 0.09016623347997665, 'sampling/importance_sampling_ratio/mean': 1.0086053609848022, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8061425983905792, 'clip_ratio/low_mean': 0.15841219387948513, 'clip_ratio/low_min': 0.15841219387948513, 'clip_ratio/high_mean': 0.08339367993175983, 'clip_ratio/high_max': 0.08339367993175983, 'clip_ratio/region_mean': 0.24180587381124496, 'reward_total_mean': 0.11938978731632233, 'reward_meter_mean': 0.30991411209106445, 'reward_meter_std': 0.31503260135650635, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.951915979385376, 'reward_repeat_penalty_std': 0.04889478534460068, 'reward_near_duplicate_penalty_mean': 0.9737776517868042, 'reward_near_duplicate_penalty_std': 0.0162584837526083, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9773937463760376, 'reward_distinct_2_std': 0.04370811954140663, 'reward_judge_quality_mean': 0.42499998211860657, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.11938978731632233, 'reward_total_composite_std': 0.13373103737831116, 'epoch': 0.0}

  1%|          | 21/3300 [03:46<9:34:29, 10.51s/it]INFO 04-12 10:49:58 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 22/3300 [03:54<8:57:15,  9.83s/it]2026-04-12 10:50:06,346 | INFO | train_grpo_train | metrics_logged mode=train step=22

                                                   
{'loss': 0.2068, 'grad_norm': 20.567657470703125, 'learning_rate': 9.936363636363638e-06, 'num_tokens': 51184.0, 'completions/mean_length': 38.375, 'completions/min_length': 32.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.375, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.8363314867019653, 'rewards/meter/std': 0.3087863326072693, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7437499761581421, 'rewards/judge_quality/std': 0.2432481348514557, 'rewards/repeat_penalty/mean': 0.9796510338783264, 'rewards/repeat_penalty/std': 0.023779788985848427, 'rewards/near_duplicate_penalty/mean': 0.9888085126876831, 'rewards/near_duplicate_penalty/std': 0.011351306922733784, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9908424615859985, 'rewards/distinct_2/std': 0.025901345536112785, 'rewards/total_composite/mean': 0.6647837162017822, 'rewards/total_composite/std': 0.3457593619823456, 'reward': 0.6647837162017822, 'reward_std': 0.3457593619823456, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10561378300189972, 'sampling/sampling_logp_difference/max': 1.8625006675720215, 'sampling/importance_sampling_ratio/min': 0.1552838236093521, 'sampling/importance_sampling_ratio/mean': 1.0084288120269775, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.727835550904274, 'clip_ratio/low_mean': 0.06501548085361719, 'clip_ratio/low_min': 0.06501548085361719, 'clip_ratio/high_mean': 0.030665107304230332, 'clip_ratio/high_max': 0.030665107304230332, 'clip_ratio/region_mean': 0.09568058815784752, 'reward_total_mean': 0.6647837162017822, 'reward_meter_mean': 0.8363314867019653, 'reward_meter_std': 0.3087863326072693, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9796510338783264, 'reward_repeat_penalty_std': 0.023779788985848427, 'reward_near_duplicate_penalty_mean': 0.9888085126876831, 'reward_near_duplicate_penalty_std': 0.011351306922733784, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9908424615859985, 'reward_distinct_2_std': 0.025901345536112785, 'reward_judge_quality_mean': 0.7437499761581421, 'reward_judge_quality_std': 0.2432481348514557, 'reward_total_composite_mean': 0.6647837162017822, 'reward_total_composite_std': 0.3457593619823456, 'epoch': 0.0}

  1%|          | 22/3300 [03:54<8:57:15,  9.83s/it]INFO 04-12 10:50:06 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 23/3300 [04:09<10:21:22, 11.38s/it]2026-04-12 10:50:21,321 | INFO | train_grpo_train | metrics_logged mode=train step=23

                                                    
{'loss': -0.0034, 'grad_norm': 2.6942813396453857, 'learning_rate': 9.933333333333334e-06, 'num_tokens': 53733.0, 'completions/mean_length': 498.625, 'completions/min_length': 405.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.75, 'completions/mean_terminated_length': 458.5, 'completions/min_terminated_length': 405.0, 'completions/max_terminated_length': 512.0, 'rewards/meter/mean': 0.8259201049804688, 'rewards/meter/std': 0.15692374110221863, 'rewards/count_adherence/mean': 0.9642857313156128, 'rewards/count_adherence/std': 0.05399492755532265, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9857219457626343, 'rewards/lexical_plausibility/std': 0.026594150811433792, 'rewards/judge_quality/mean': 0.20000000298023224, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.9891016483306885, 'rewards/repeat_penalty/std': 0.008064168505370617, 'rewards/near_duplicate_penalty/mean': 0.9955102801322937, 'rewards/near_duplicate_penalty/std': 0.0016857460141181946, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9935599565505981, 'rewards/distinct_2/std': 0.0075568766333162785, 'rewards/total_composite/mean': 0.1499270498752594, 'rewards/total_composite/std': 0.042359739542007446, 'reward': 0.1499270498752594, 'reward_std': 0.042359739542007446, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2144458144903183, 'sampling/sampling_logp_difference/max': 1.8869943618774414, 'sampling/importance_sampling_ratio/min': 0.1515265703201294, 'sampling/importance_sampling_ratio/mean': 1.034726619720459, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.765936017036438, 'clip_ratio/low_mean': 0.021484375, 'clip_ratio/low_min': 0.021484375, 'clip_ratio/high_mean': 0.024074073880910873, 'clip_ratio/high_max': 0.024074073880910873, 'clip_ratio/region_mean': 0.04555844888091087, 'reward_total_mean': 0.1499270498752594, 'reward_meter_mean': 0.8259201049804688, 'reward_meter_std': 0.15692374110221863, 'reward_count_adherence_mean': 0.9642857313156128, 'reward_count_adherence_std': 0.05399492755532265, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9857219457626343, 'reward_lexical_plausibility_std': 0.026594150811433792, 'reward_repeat_penalty_mean': 0.9891016483306885, 'reward_repeat_penalty_std': 0.008064168505370617, 'reward_near_duplicate_penalty_mean': 0.9955102801322937, 'reward_near_duplicate_penalty_std': 0.0016857460141181946, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9935599565505981, 'reward_distinct_2_std': 0.0075568766333162785, 'reward_judge_quality_mean': 0.20000000298023224, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.1499270498752594, 'reward_total_composite_std': 0.042359739542007446, 'epoch': 0.0}

  1%|          | 23/3300 [04:09<10:21:22, 11.38s/it]INFO 04-12 10:50:21 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 24/3300 [04:19<10:00:07, 10.99s/it]2026-04-12 10:50:31,413 | INFO | train_grpo_train | metrics_logged mode=train step=24

                                                    
{'loss': 0.0694, 'grad_norm': 11.913330078125, 'learning_rate': 9.930303030303031e-06, 'num_tokens': 55755.0, 'completions/mean_length': 85.75, 'completions/min_length': 53.0, 'completions/max_length': 124.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 85.75, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 124.0, 'rewards/meter/mean': 0.516542911529541, 'rewards/meter/std': 0.34262797236442566, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.36249998211860657, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.9855961203575134, 'rewards/repeat_penalty/std': 0.01391402818262577, 'rewards/near_duplicate_penalty/mean': 0.9855961203575134, 'rewards/near_duplicate_penalty/std': 0.01391402818262577, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.14416193962097168, 'rewards/total_composite/std': 0.11695463955402374, 'reward': 0.14416193962097168, 'reward_std': 0.11695463955402374, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21823984384536743, 'sampling/sampling_logp_difference/max': 2.2229785919189453, 'sampling/importance_sampling_ratio/min': 0.10828609019517899, 'sampling/importance_sampling_ratio/mean': 1.012178897857666, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1747989654541016, 'clip_ratio/low_mean': 0.07377693615853786, 'clip_ratio/low_min': 0.07377693615853786, 'clip_ratio/high_mean': 0.15401890873908997, 'clip_ratio/high_max': 0.15401890873908997, 'clip_ratio/region_mean': 0.22779584489762783, 'reward_total_mean': 0.14416193962097168, 'reward_meter_mean': 0.516542911529541, 'reward_meter_std': 0.34262797236442566, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9855961203575134, 'reward_repeat_penalty_std': 0.01391402818262577, 'reward_near_duplicate_penalty_mean': 0.9855961203575134, 'reward_near_duplicate_penalty_std': 0.01391402818262577, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.36249998211860657, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.14416193962097168, 'reward_total_composite_std': 0.11695463955402374, 'epoch': 0.0}

  1%|          | 24/3300 [04:19<10:00:07, 10.99s/it]INFO 04-12 10:50:31 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 25/3300 [04:28<9:32:33, 10.49s/it] 2026-04-12 10:50:40,732 | INFO | train_grpo_train | metrics_logged mode=train step=25

                                                   
{'loss': 0.1321, 'grad_norm': 11.314960479736328, 'learning_rate': 9.927272727272728e-06, 'num_tokens': 57941.0, 'completions/mean_length': 88.25, 'completions/min_length': 69.0, 'completions/max_length': 111.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 88.25, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 111.0, 'rewards/meter/mean': 0.3611519932746887, 'rewards/meter/std': 0.3820111155509949, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9807378053665161, 'rewards/lexical_plausibility/std': 0.029874874278903008, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.0353553369641304, 'rewards/repeat_penalty/mean': 0.9859764575958252, 'rewards/repeat_penalty/std': 0.022083619609475136, 'rewards/near_duplicate_penalty/mean': 0.992519736289978, 'rewards/near_duplicate_penalty/std': 0.007646000944077969, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9933686852455139, 'rewards/distinct_2/std': 0.01875614933669567, 'rewards/total_composite/mean': 0.13510285317897797, 'rewards/total_composite/std': 0.14426524937152863, 'reward': 0.13510285317897797, 'reward_std': 0.14426523447036743, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1977376788854599, 'sampling/sampling_logp_difference/max': 1.5759835243225098, 'sampling/importance_sampling_ratio/min': 0.2068040668964386, 'sampling/importance_sampling_ratio/mean': 1.0206221342086792, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7679535001516342, 'clip_ratio/low_mean': 0.11564315110445023, 'clip_ratio/low_min': 0.11564315110445023, 'clip_ratio/high_mean': 0.07766031473875046, 'clip_ratio/high_max': 0.07766031473875046, 'clip_ratio/region_mean': 0.19330346584320068, 'reward_total_mean': 0.13510285317897797, 'reward_meter_mean': 0.3611519932746887, 'reward_meter_std': 0.3820111155509949, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9807378053665161, 'reward_lexical_plausibility_std': 0.029874874278903008, 'reward_repeat_penalty_mean': 0.9859764575958252, 'reward_repeat_penalty_std': 0.022083619609475136, 'reward_near_duplicate_penalty_mean': 0.992519736289978, 'reward_near_duplicate_penalty_std': 0.007646000944077969, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9933686852455139, 'reward_distinct_2_std': 0.01875614933669567, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.0353553369641304, 'reward_total_composite_mean': 0.13510285317897797, 'reward_total_composite_std': 0.14426524937152863, 'epoch': 0.0}

  1%|          | 25/3300 [04:28<9:32:33, 10.49s/it]INFO 04-12 10:50:40 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 26/3300 [04:37<9:11:00, 10.10s/it]2026-04-12 10:50:49,918 | INFO | train_grpo_train | metrics_logged mode=train step=26

                                                   
{'loss': 0.0648, 'grad_norm': 16.19539451599121, 'learning_rate': 9.924242424242425e-06, 'num_tokens': 59731.0, 'completions/mean_length': 64.75, 'completions/min_length': 46.0, 'completions/max_length': 82.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.75, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.5833607316017151, 'rewards/meter/std': 0.4440600275993347, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9788110256195068, 'rewards/lexical_plausibility/std': 0.05993138626217842, 'rewards/judge_quality/mean': 0.5174999833106995, 'rewards/judge_quality/std': 0.2522329092025757, 'rewards/repeat_penalty/mean': 0.9924741387367249, 'rewards/repeat_penalty/std': 0.011241739615797997, 'rewards/near_duplicate_penalty/mean': 0.9924741387367249, 'rewards/near_duplicate_penalty/std': 0.011241739615797997, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2691909372806549, 'rewards/total_composite/std': 0.2605314254760742, 'reward': 0.2691909372806549, 'reward_std': 0.2605314254760742, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2431088238954544, 'sampling/sampling_logp_difference/max': 1.6622037887573242, 'sampling/importance_sampling_ratio/min': 0.18972042202949524, 'sampling/importance_sampling_ratio/mean': 1.0399829149246216, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0710978358983994, 'clip_ratio/low_mean': 0.08128589391708374, 'clip_ratio/low_min': 0.08128589391708374, 'clip_ratio/high_mean': 0.1346008237451315, 'clip_ratio/high_max': 0.1346008237451315, 'clip_ratio/region_mean': 0.21588671766221523, 'reward_total_mean': 0.2691909372806549, 'reward_meter_mean': 0.5833607316017151, 'reward_meter_std': 0.4440600275993347, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9788110256195068, 'reward_lexical_plausibility_std': 0.05993138626217842, 'reward_repeat_penalty_mean': 0.9924741387367249, 'reward_repeat_penalty_std': 0.011241739615797997, 'reward_near_duplicate_penalty_mean': 0.9924741387367249, 'reward_near_duplicate_penalty_std': 0.011241739615797997, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5174999833106995, 'reward_judge_quality_std': 0.2522329092025757, 'reward_total_composite_mean': 0.2691909372806549, 'reward_total_composite_std': 0.2605314254760742, 'epoch': 0.0}

  1%|          | 26/3300 [04:37<9:11:00, 10.10s/it]INFO 04-12 10:50:50 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 27/3300 [04:47<8:59:45,  9.89s/it]2026-04-12 10:50:59,337 | INFO | train_grpo_train | metrics_logged mode=train step=27

                                                   
{'loss': 0.054, 'grad_norm': 18.46763801574707, 'learning_rate': 9.921212121212121e-06, 'num_tokens': 61588.0, 'completions/mean_length': 57.125, 'completions/min_length': 42.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.125, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.5880099534988403, 'rewards/meter/std': 0.3607083559036255, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/repeat_penalty/mean': 0.9886506795883179, 'rewards/repeat_penalty/std': 0.02050127275288105, 'rewards/near_duplicate_penalty/mean': 0.9960471391677856, 'rewards/near_duplicate_penalty/std': 0.007033475674688816, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9926035404205322, 'rewards/distinct_2/std': 0.020920326933264732, 'rewards/total_composite/mean': 0.18190374970436096, 'rewards/total_composite/std': 0.11817481368780136, 'reward': 0.18190374970436096, 'reward_std': 0.11817480623722076, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23327620327472687, 'sampling/sampling_logp_difference/max': 1.5300722122192383, 'sampling/importance_sampling_ratio/min': 0.2165200412273407, 'sampling/importance_sampling_ratio/mean': 1.0309526920318604, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9116510599851608, 'clip_ratio/low_mean': 0.11785485781729221, 'clip_ratio/low_min': 0.11785485781729221, 'clip_ratio/high_mean': 0.0773110780864954, 'clip_ratio/high_max': 0.0773110780864954, 'clip_ratio/region_mean': 0.1951659359037876, 'reward_total_mean': 0.18190374970436096, 'reward_meter_mean': 0.5880099534988403, 'reward_meter_std': 0.3607083559036255, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9886506795883179, 'reward_repeat_penalty_std': 0.02050127275288105, 'reward_near_duplicate_penalty_mean': 0.9960471391677856, 'reward_near_duplicate_penalty_std': 0.007033475674688816, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9926035404205322, 'reward_distinct_2_std': 0.020920326933264732, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.18190374970436096, 'reward_total_composite_std': 0.11817481368780136, 'epoch': 0.0}

  1%|          | 27/3300 [04:47<8:59:45,  9.89s/it]INFO 04-12 10:50:59 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 28/3300 [04:55<8:35:28,  9.45s/it]2026-04-12 10:51:07,758 | INFO | train_grpo_train | metrics_logged mode=train step=28

                                                   
{'loss': -0.0355, 'grad_norm': 16.11900520324707, 'learning_rate': 9.918181818181818e-06, 'num_tokens': 63174.0, 'completions/mean_length': 51.25, 'completions/min_length': 37.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.25, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.7667711973190308, 'rewards/meter/std': 0.35821136832237244, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9875565767288208, 'rewards/lexical_plausibility/std': 0.025101033970713615, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.9867918491363525, 'rewards/repeat_penalty/std': 0.028985347598791122, 'rewards/near_duplicate_penalty/mean': 0.9867918491363525, 'rewards/near_duplicate_penalty/std': 0.028985347598791122, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2989039719104767, 'rewards/total_composite/std': 0.16186337172985077, 'reward': 0.2989039719104767, 'reward_std': 0.16186335682868958, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2023225873708725, 'sampling/sampling_logp_difference/max': 1.1405749320983887, 'sampling/importance_sampling_ratio/min': 0.31963521242141724, 'sampling/importance_sampling_ratio/mean': 1.0361260175704956, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1601669788360596, 'clip_ratio/low_mean': 0.041732910089194775, 'clip_ratio/low_min': 0.041732910089194775, 'clip_ratio/high_mean': 0.1668364442884922, 'clip_ratio/high_max': 0.1668364442884922, 'clip_ratio/region_mean': 0.20856935437768698, 'reward_total_mean': 0.2989039719104767, 'reward_meter_mean': 0.7667711973190308, 'reward_meter_std': 0.35821136832237244, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9875565767288208, 'reward_lexical_plausibility_std': 0.025101033970713615, 'reward_repeat_penalty_mean': 0.9867918491363525, 'reward_repeat_penalty_std': 0.028985347598791122, 'reward_near_duplicate_penalty_mean': 0.9867918491363525, 'reward_near_duplicate_penalty_std': 0.028985347598791122, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.2989039719104767, 'reward_total_composite_std': 0.16186337172985077, 'epoch': 0.0}

  1%|          | 28/3300 [04:55<8:35:28,  9.45s/it]INFO 04-12 10:51:07 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 29/3300 [05:09<9:52:56, 10.88s/it]2026-04-12 10:51:21,955 | INFO | train_grpo_train | metrics_logged mode=train step=29

                                                   
{'loss': 0.109, 'grad_norm': 8.28359603881836, 'learning_rate': 9.915151515151515e-06, 'num_tokens': 66740.0, 'completions/mean_length': 208.75, 'completions/min_length': 181.0, 'completions/max_length': 240.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 208.75, 'completions/min_terminated_length': 181.0, 'completions/max_terminated_length': 240.0, 'rewards/meter/mean': 0.1128695011138916, 'rewards/meter/std': 0.04446408152580261, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.0578637570142746, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.9927715063095093, 'rewards/repeat_penalty/std': 0.007351625245064497, 'rewards/near_duplicate_penalty/mean': 0.9953268766403198, 'rewards/near_duplicate_penalty/std': 0.003063744166865945, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9974359273910522, 'rewards/distinct_2/std': 0.007252376992255449, 'rewards/total_composite/mean': 0.03924858197569847, 'rewards/total_composite/std': 0.014390570111572742, 'reward': 0.03924858197569847, 'reward_std': 0.014390568248927593, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23383566737174988, 'sampling/sampling_logp_difference/max': 2.1666746139526367, 'sampling/importance_sampling_ratio/min': 0.11455792933702469, 'sampling/importance_sampling_ratio/mean': 1.0412983894348145, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.5335609018802643, 'clip_ratio/low_mean': 0.10280177742242813, 'clip_ratio/low_min': 0.10280177742242813, 'clip_ratio/high_mean': 0.11149711534380913, 'clip_ratio/high_max': 0.11149711534380913, 'clip_ratio/region_mean': 0.21429889276623726, 'reward_total_mean': 0.03924858197569847, 'reward_meter_mean': 0.1128695011138916, 'reward_meter_std': 0.04446408152580261, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.0578637570142746, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9927715063095093, 'reward_repeat_penalty_std': 0.007351625245064497, 'reward_near_duplicate_penalty_mean': 0.9953268766403198, 'reward_near_duplicate_penalty_std': 0.003063744166865945, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9974359273910522, 'reward_distinct_2_std': 0.007252376992255449, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.03924858197569847, 'reward_total_composite_std': 0.014390570111572742, 'epoch': 0.0}

  1%|          | 29/3300 [05:09<9:52:56, 10.88s/it]INFO 04-12 10:51:22 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 30/3300 [05:19<9:38:47, 10.62s/it]2026-04-12 10:51:31,977 | INFO | train_grpo_train | metrics_logged mode=train step=30

                                                   
{'loss': 0.0386, 'grad_norm': 11.373634338378906, 'learning_rate': 9.912121212121213e-06, 'num_tokens': 68703.0, 'completions/mean_length': 81.375, 'completions/min_length': 65.0, 'completions/max_length': 94.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 81.375, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.20140917599201202, 'rewards/meter/std': 0.25694817304611206, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.48374998569488525, 'rewards/judge_quality/std': 0.18196842074394226, 'rewards/repeat_penalty/mean': 0.9899923205375671, 'rewards/repeat_penalty/std': 0.008522940799593925, 'rewards/near_duplicate_penalty/mean': 0.9899923205375671, 'rewards/near_duplicate_penalty/std': 0.008522940799593925, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.08122860640287399, 'rewards/total_composite/std': 0.09018824994564056, 'reward': 0.08122860640287399, 'reward_std': 0.09018824994564056, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20405979454517365, 'sampling/sampling_logp_difference/max': 1.3296279907226562, 'sampling/importance_sampling_ratio/min': 0.26457566022872925, 'sampling/importance_sampling_ratio/mean': 1.0039526224136353, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7351030707359314, 'clip_ratio/low_mean': 0.14230559766292572, 'clip_ratio/low_min': 0.14230559766292572, 'clip_ratio/high_mean': 0.050825824961066246, 'clip_ratio/high_max': 0.050825824961066246, 'clip_ratio/region_mean': 0.19313142262399197, 'reward_total_mean': 0.08122860640287399, 'reward_meter_mean': 0.20140917599201202, 'reward_meter_std': 0.25694817304611206, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9899923205375671, 'reward_repeat_penalty_std': 0.008522940799593925, 'reward_near_duplicate_penalty_mean': 0.9899923205375671, 'reward_near_duplicate_penalty_std': 0.008522940799593925, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.48374998569488525, 'reward_judge_quality_std': 0.18196842074394226, 'reward_total_composite_mean': 0.08122860640287399, 'reward_total_composite_std': 0.09018824994564056, 'epoch': 0.0}

  1%|          | 30/3300 [05:19<9:38:47, 10.62s/it]INFO 04-12 10:51:32 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 31/3300 [05:29<9:20:43, 10.29s/it]2026-04-12 10:51:41,504 | INFO | train_grpo_train | metrics_logged mode=train step=31

                                                   
{'loss': 0.0581, 'grad_norm': 13.426535606384277, 'learning_rate': 9.90909090909091e-06, 'num_tokens': 70320.0, 'completions/mean_length': 52.125, 'completions/min_length': 34.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.125, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.6859865188598633, 'rewards/meter/std': 0.33277973532676697, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5174999833106995, 'rewards/judge_quality/std': 0.25783440470695496, 'rewards/repeat_penalty/mean': 0.9892197847366333, 'rewards/repeat_penalty/std': 0.0152900917455554, 'rewards/near_duplicate_penalty/mean': 0.9892197847366333, 'rewards/near_duplicate_penalty/std': 0.0152900917455554, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3664495348930359, 'rewards/total_composite/std': 0.3514110743999481, 'reward': 0.3664495348930359, 'reward_std': 0.3514110743999481, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21009325981140137, 'sampling/sampling_logp_difference/max': 2.2127251625061035, 'sampling/importance_sampling_ratio/min': 0.10940210521221161, 'sampling/importance_sampling_ratio/mean': 1.0185762643814087, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6987058073282242, 'clip_ratio/low_mean': 0.0922289052978158, 'clip_ratio/low_min': 0.0922289052978158, 'clip_ratio/high_mean': 0.07851307280361652, 'clip_ratio/high_max': 0.07851307280361652, 'clip_ratio/region_mean': 0.17074197810143232, 'reward_total_mean': 0.3664495348930359, 'reward_meter_mean': 0.6859865188598633, 'reward_meter_std': 0.33277973532676697, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9892197847366333, 'reward_repeat_penalty_std': 0.0152900917455554, 'reward_near_duplicate_penalty_mean': 0.9892197847366333, 'reward_near_duplicate_penalty_std': 0.0152900917455554, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5174999833106995, 'reward_judge_quality_std': 0.25783440470695496, 'reward_total_composite_mean': 0.3664495348930359, 'reward_total_composite_std': 0.3514110743999481, 'epoch': 0.0}

  1%|          | 31/3300 [05:29<9:20:43, 10.29s/it]INFO 04-12 10:51:41 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 32/3300 [05:38<9:03:25,  9.98s/it]2026-04-12 10:51:50,747 | INFO | train_grpo_train | metrics_logged mode=train step=32

                                                   
{'loss': 0.0553, 'grad_norm': 13.0629243850708, 'learning_rate': 9.906060606060607e-06, 'num_tokens': 71997.0, 'completions/mean_length': 47.625, 'completions/min_length': 32.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.625, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.6337600946426392, 'rewards/meter/std': 0.4233327805995941, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.36249998211860657, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.9900386929512024, 'rewards/repeat_penalty/std': 0.01413870882242918, 'rewards/near_duplicate_penalty/mean': 0.9900386929512024, 'rewards/near_duplicate_penalty/std': 0.01413870882242918, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.20086389780044556, 'rewards/total_composite/std': 0.18254370987415314, 'reward': 0.20086389780044556, 'reward_std': 0.18254370987415314, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19689208269119263, 'sampling/sampling_logp_difference/max': 2.0852737426757812, 'sampling/importance_sampling_ratio/min': 0.12427309155464172, 'sampling/importance_sampling_ratio/mean': 1.0180888175964355, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.149626463651657, 'clip_ratio/low_mean': 0.09710591286420822, 'clip_ratio/low_min': 0.09710591286420822, 'clip_ratio/high_mean': 0.12454579956829548, 'clip_ratio/high_max': 0.12454579956829548, 'clip_ratio/region_mean': 0.2216517124325037, 'reward_total_mean': 0.20086389780044556, 'reward_meter_mean': 0.6337600946426392, 'reward_meter_std': 0.4233327805995941, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9900386929512024, 'reward_repeat_penalty_std': 0.01413870882242918, 'reward_near_duplicate_penalty_mean': 0.9900386929512024, 'reward_near_duplicate_penalty_std': 0.01413870882242918, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.36249998211860657, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.20086389780044556, 'reward_total_composite_std': 0.18254370987415314, 'epoch': 0.0}

  1%|          | 32/3300 [05:38<9:03:25,  9.98s/it]INFO 04-12 10:51:50 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 33/3300 [05:49<9:21:05, 10.30s/it]2026-04-12 10:52:01,817 | INFO | train_grpo_train | metrics_logged mode=train step=33

                                                   
{'loss': 0.0425, 'grad_norm': 7.572989463806152, 'learning_rate': 9.903030303030305e-06, 'num_tokens': 74906.0, 'completions/mean_length': 174.625, 'completions/min_length': 118.0, 'completions/max_length': 226.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 174.625, 'completions/min_terminated_length': 118.0, 'completions/max_terminated_length': 226.0, 'rewards/meter/mean': 0.6810822486877441, 'rewards/meter/std': 0.2588964104652405, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.08625820279121399, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9818412065505981, 'rewards/lexical_plausibility/std': 0.028247511014342308, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.9900496602058411, 'rewards/repeat_penalty/std': 0.013502294197678566, 'rewards/near_duplicate_penalty/mean': 0.996315598487854, 'rewards/near_duplicate_penalty/std': 0.002750101499259472, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9936931133270264, 'rewards/distinct_2/std': 0.011680135503411293, 'rewards/total_composite/mean': 0.178521990776062, 'rewards/total_composite/std': 0.07563900202512741, 'reward': 0.178521990776062, 'reward_std': 0.07563900202512741, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2178238481283188, 'sampling/sampling_logp_difference/max': 2.948610305786133, 'sampling/importance_sampling_ratio/min': 0.0524124950170517, 'sampling/importance_sampling_ratio/mean': 1.042715311050415, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6625306010246277, 'clip_ratio/low_mean': 0.12989515997469425, 'clip_ratio/low_min': 0.12989515997469425, 'clip_ratio/high_mean': 0.07445338554680347, 'clip_ratio/high_max': 0.07445338554680347, 'clip_ratio/region_mean': 0.20434854552149773, 'reward_total_mean': 0.178521990776062, 'reward_meter_mean': 0.6810822486877441, 'reward_meter_std': 0.2588964104652405, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.08625820279121399, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9818412065505981, 'reward_lexical_plausibility_std': 0.028247511014342308, 'reward_repeat_penalty_mean': 0.9900496602058411, 'reward_repeat_penalty_std': 0.013502294197678566, 'reward_near_duplicate_penalty_mean': 0.996315598487854, 'reward_near_duplicate_penalty_std': 0.002750101499259472, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9936931133270264, 'reward_distinct_2_std': 0.011680135503411293, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.178521990776062, 'reward_total_composite_std': 0.07563900202512741, 'epoch': 0.0}

  1%|          | 33/3300 [05:49<9:21:05, 10.30s/it]INFO 04-12 10:52:02 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 34/3300 [06:00<9:29:02, 10.45s/it]2026-04-12 10:52:12,617 | INFO | train_grpo_train | metrics_logged mode=train step=34

                                                   
{'loss': 0.0509, 'grad_norm': 9.460807800292969, 'learning_rate': 9.9e-06, 'num_tokens': 77072.0, 'completions/mean_length': 105.75, 'completions/min_length': 74.0, 'completions/max_length': 133.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 105.75, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 133.0, 'rewards/meter/mean': 0.49930068850517273, 'rewards/meter/std': 0.4238707423210144, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.995398998260498, 'rewards/lexical_plausibility/std': 0.013013492338359356, 'rewards/judge_quality/mean': 0.4375, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.9959745407104492, 'rewards/repeat_penalty/std': 0.0037175368051975965, 'rewards/near_duplicate_penalty/mean': 0.9959745407104492, 'rewards/near_duplicate_penalty/std': 0.0037175368051975965, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.24013672769069672, 'rewards/total_composite/std': 0.23535257577896118, 'reward': 0.24013672769069672, 'reward_std': 0.23535257577896118, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20552921295166016, 'sampling/sampling_logp_difference/max': 1.8478271961212158, 'sampling/importance_sampling_ratio/min': 0.1575791835784912, 'sampling/importance_sampling_ratio/mean': 1.0442026853561401, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.3008000552654266, 'clip_ratio/low_mean': 0.12665731459856033, 'clip_ratio/low_min': 0.12665731459856033, 'clip_ratio/high_mean': 0.08025272376835346, 'clip_ratio/high_max': 0.08025272376835346, 'clip_ratio/region_mean': 0.2069100383669138, 'reward_total_mean': 0.24013672769069672, 'reward_meter_mean': 0.49930068850517273, 'reward_meter_std': 0.4238707423210144, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.995398998260498, 'reward_lexical_plausibility_std': 0.013013492338359356, 'reward_repeat_penalty_mean': 0.9959745407104492, 'reward_repeat_penalty_std': 0.0037175368051975965, 'reward_near_duplicate_penalty_mean': 0.9959745407104492, 'reward_near_duplicate_penalty_std': 0.0037175368051975965, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4375, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.24013672769069672, 'reward_total_composite_std': 0.23535257577896118, 'epoch': 0.0}

  1%|          | 34/3300 [06:00<9:29:02, 10.45s/it]INFO 04-12 10:52:12 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 35/3300 [06:13<10:01:29, 11.05s/it]2026-04-12 10:52:25,070 | INFO | train_grpo_train | metrics_logged mode=train step=35

                                                    
{'loss': 0.033, 'grad_norm': 5.8426289558410645, 'learning_rate': 9.896969696969699e-06, 'num_tokens': 81030.0, 'completions/mean_length': 284.75, 'completions/min_length': 189.0, 'completions/max_length': 331.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 284.75, 'completions/min_terminated_length': 189.0, 'completions/max_terminated_length': 331.0, 'rewards/meter/mean': 0.46129634976387024, 'rewards/meter/std': 0.23356977105140686, 'rewards/count_adherence/mean': 0.9027777910232544, 'rewards/count_adherence/std': 0.07120776921510696, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9986007213592529, 'rewards/lexical_plausibility/std': 0.003957694862037897, 'rewards/judge_quality/mean': 0.20000000298023224, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.9915546178817749, 'rewards/repeat_penalty/std': 0.007529374677687883, 'rewards/near_duplicate_penalty/mean': 0.9934437870979309, 'rewards/near_duplicate_penalty/std': 0.004776338115334511, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9980959892272949, 'rewards/distinct_2/std': 0.005385419353842735, 'rewards/total_composite/mean': 0.06845705211162567, 'rewards/total_composite/std': 0.058959536254405975, 'reward': 0.06845705211162567, 'reward_std': 0.05895953252911568, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21234013140201569, 'sampling/sampling_logp_difference/max': 2.172140121459961, 'sampling/importance_sampling_ratio/min': 0.11393352597951889, 'sampling/importance_sampling_ratio/mean': 1.0409951210021973, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.7264479994773865, 'clip_ratio/low_mean': 0.07352687604725361, 'clip_ratio/low_min': 0.07352687604725361, 'clip_ratio/high_mean': 0.13289386592805386, 'clip_ratio/high_max': 0.13289386592805386, 'clip_ratio/region_mean': 0.20642074197530746, 'reward_total_mean': 0.06845705211162567, 'reward_meter_mean': 0.46129634976387024, 'reward_meter_std': 0.23356977105140686, 'reward_count_adherence_mean': 0.9027777910232544, 'reward_count_adherence_std': 0.07120776921510696, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9986007213592529, 'reward_lexical_plausibility_std': 0.003957694862037897, 'reward_repeat_penalty_mean': 0.9915546178817749, 'reward_repeat_penalty_std': 0.007529374677687883, 'reward_near_duplicate_penalty_mean': 0.9934437870979309, 'reward_near_duplicate_penalty_std': 0.004776338115334511, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9980959892272949, 'reward_distinct_2_std': 0.005385419353842735, 'reward_judge_quality_mean': 0.20000000298023224, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.06845705211162567, 'reward_total_composite_std': 0.058959536254405975, 'epoch': 0.0}

  1%|          | 35/3300 [06:13<10:01:29, 11.05s/it]INFO 04-12 10:52:25 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 36/3300 [06:28<11:06:51, 12.26s/it]2026-04-12 10:52:40,140 | INFO | train_grpo_train | metrics_logged mode=train step=36

                                                    
{'loss': -0.0778, 'grad_norm': 4.092381477355957, 'learning_rate': 9.893939393939395e-06, 'num_tokens': 84863.0, 'completions/mean_length': 449.125, 'completions/min_length': 310.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 411.3999938964844, 'completions/min_terminated_length': 310.0, 'completions/max_terminated_length': 464.0, 'rewards/meter/mean': 0.1800491064786911, 'rewards/meter/std': 0.09611400961875916, 'rewards/count_adherence/mean': 0.9117647409439087, 'rewards/count_adherence/std': 0.06288499385118484, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.20000000298023224, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.9814121723175049, 'rewards/repeat_penalty/std': 0.020063040778040886, 'rewards/near_duplicate_penalty/mean': 0.9937238693237305, 'rewards/near_duplicate_penalty/std': 0.005176292732357979, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9875483512878418, 'rewards/distinct_2/std': 0.015593364834785461, 'rewards/total_composite/mean': 0.03374875336885452, 'rewards/total_composite/std': 0.03367682546377182, 'reward': 0.03374875336885452, 'reward_std': 0.03367682173848152, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21900925040245056, 'sampling/sampling_logp_difference/max': 2.311619758605957, 'sampling/importance_sampling_ratio/min': 0.09910060465335846, 'sampling/importance_sampling_ratio/mean': 1.0577551126480103, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0560347735881805, 'clip_ratio/low_mean': 0.06574058718979359, 'clip_ratio/low_min': 0.06574058718979359, 'clip_ratio/high_mean': 0.04515678249299526, 'clip_ratio/high_max': 0.04515678249299526, 'clip_ratio/region_mean': 0.11089736968278885, 'reward_total_mean': 0.03374875336885452, 'reward_meter_mean': 0.1800491064786911, 'reward_meter_std': 0.09611400961875916, 'reward_count_adherence_mean': 0.9117647409439087, 'reward_count_adherence_std': 0.06288499385118484, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9814121723175049, 'reward_repeat_penalty_std': 0.020063040778040886, 'reward_near_duplicate_penalty_mean': 0.9937238693237305, 'reward_near_duplicate_penalty_std': 0.005176292732357979, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9875483512878418, 'reward_distinct_2_std': 0.015593364834785461, 'reward_judge_quality_mean': 0.20000000298023224, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.03374875336885452, 'reward_total_composite_std': 0.03367682546377182, 'epoch': 0.0}

  1%|          | 36/3300 [06:28<11:06:51, 12.26s/it]INFO 04-12 10:52:40 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 37/3300 [06:38<10:32:39, 11.63s/it]2026-04-12 10:52:50,316 | INFO | train_grpo_train | metrics_logged mode=train step=37

                                                    
{'loss': 0.1322, 'grad_norm': 11.867147445678711, 'learning_rate': 9.890909090909092e-06, 'num_tokens': 86734.0, 'completions/mean_length': 57.875, 'completions/min_length': 42.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.875, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.9086538553237915, 'rewards/meter/std': 0.1690373420715332, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.42499998211860657, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.9959057569503784, 'rewards/repeat_penalty/std': 0.011237816885113716, 'rewards/near_duplicate_penalty/mean': 0.9959057569503784, 'rewards/near_duplicate_penalty/std': 0.011237816885113716, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4023837745189667, 'rewards/total_composite/std': 0.15859727561473846, 'reward': 0.4023837745189667, 'reward_std': 0.15859729051589966, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1997494101524353, 'sampling/sampling_logp_difference/max': 1.7911901473999023, 'sampling/importance_sampling_ratio/min': 0.1667615920305252, 'sampling/importance_sampling_ratio/mean': 1.0365508794784546, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.137260004878044, 'clip_ratio/low_mean': 0.05545040871948004, 'clip_ratio/low_min': 0.05545040871948004, 'clip_ratio/high_mean': 0.11482405569404364, 'clip_ratio/high_max': 0.11482405569404364, 'clip_ratio/region_mean': 0.17027446441352367, 'reward_total_mean': 0.4023837745189667, 'reward_meter_mean': 0.9086538553237915, 'reward_meter_std': 0.1690373420715332, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9959057569503784, 'reward_repeat_penalty_std': 0.011237816885113716, 'reward_near_duplicate_penalty_mean': 0.9959057569503784, 'reward_near_duplicate_penalty_std': 0.011237816885113716, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.42499998211860657, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.4023837745189667, 'reward_total_composite_std': 0.15859727561473846, 'epoch': 0.0}

  1%|          | 37/3300 [06:38<10:32:39, 11.63s/it]INFO 04-12 10:52:50 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 38/3300 [06:47<9:55:55, 10.96s/it] 2026-04-12 10:52:59,710 | INFO | train_grpo_train | metrics_logged mode=train step=38

                                                   
{'loss': 0.0979, 'grad_norm': 11.908285140991211, 'learning_rate': 9.887878787878789e-06, 'num_tokens': 88228.0, 'completions/mean_length': 41.75, 'completions/min_length': 33.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.75, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.2573082745075226, 'rewards/meter/std': 0.3478483557701111, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9866071343421936, 'rewards/lexical_plausibility/std': 0.03788072615861893, 'rewards/judge_quality/mean': 0.4087499976158142, 'rewards/judge_quality/std': 0.27445724606513977, 'rewards/repeat_penalty/mean': 0.977064847946167, 'rewards/repeat_penalty/std': 0.02895965613424778, 'rewards/near_duplicate_penalty/mean': 0.977064847946167, 'rewards/near_duplicate_penalty/std': 0.02895965613424778, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.11427588760852814, 'rewards/total_composite/std': 0.21644611656665802, 'reward': 0.11427588760852814, 'reward_std': 0.21644611656665802, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19935156404972076, 'sampling/sampling_logp_difference/max': 1.3580989837646484, 'sampling/importance_sampling_ratio/min': 0.2571491599082947, 'sampling/importance_sampling_ratio/mean': 1.0438369512557983, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6971920430660248, 'clip_ratio/low_mean': 0.2066736575216055, 'clip_ratio/low_min': 0.2066736575216055, 'clip_ratio/high_mean': 0.022727273404598236, 'clip_ratio/high_max': 0.022727273404598236, 'clip_ratio/region_mean': 0.22940093092620373, 'reward_total_mean': 0.11427588760852814, 'reward_meter_mean': 0.2573082745075226, 'reward_meter_std': 0.3478483557701111, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9866071343421936, 'reward_lexical_plausibility_std': 0.03788072615861893, 'reward_repeat_penalty_mean': 0.977064847946167, 'reward_repeat_penalty_std': 0.02895965613424778, 'reward_near_duplicate_penalty_mean': 0.977064847946167, 'reward_near_duplicate_penalty_std': 0.02895965613424778, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4087499976158142, 'reward_judge_quality_std': 0.27445724606513977, 'reward_total_composite_mean': 0.11427588760852814, 'reward_total_composite_std': 0.21644611656665802, 'epoch': 0.0}

  1%|          | 38/3300 [06:47<9:55:55, 10.96s/it]INFO 04-12 10:52:59 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 39/3300 [07:02<10:52:40, 12.01s/it]2026-04-12 10:53:14,161 | INFO | train_grpo_train | metrics_logged mode=train step=39

                                                    
{'loss': 0.0351, 'grad_norm': 4.0106892585754395, 'learning_rate': 9.884848484848486e-06, 'num_tokens': 93289.0, 'completions/mean_length': 451.625, 'completions/min_length': 323.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 443.0000305175781, 'completions/min_terminated_length': 323.0, 'completions/max_terminated_length': 511.0, 'rewards/meter/mean': 0.7355791330337524, 'rewards/meter/std': 0.31982097029685974, 'rewards/count_adherence/mean': 0.942307710647583, 'rewards/count_adherence/std': 0.05439282953739166, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.17500001192092896, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/repeat_penalty/mean': 0.9885231256484985, 'rewards/repeat_penalty/std': 0.012684613466262817, 'rewards/near_duplicate_penalty/mean': 0.995418906211853, 'rewards/near_duplicate_penalty/std': 0.0017438719514757395, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9930649995803833, 'rewards/distinct_2/std': 0.01195761002600193, 'rewards/total_composite/mean': 0.07325583696365356, 'rewards/total_composite/std': 0.059081532061100006, 'reward': 0.07325583696365356, 'reward_std': 0.059081532061100006, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19850671291351318, 'sampling/sampling_logp_difference/max': 1.7185239791870117, 'sampling/importance_sampling_ratio/min': 0.17933064699172974, 'sampling/importance_sampling_ratio/mean': 1.0332156419754028, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.398359000682831, 'clip_ratio/low_mean': 0.06206880137324333, 'clip_ratio/low_min': 0.06206880137324333, 'clip_ratio/high_mean': 0.08713553473353386, 'clip_ratio/high_max': 0.08713553473353386, 'clip_ratio/region_mean': 0.1492043361067772, 'reward_total_mean': 0.07325583696365356, 'reward_meter_mean': 0.7355791330337524, 'reward_meter_std': 0.31982097029685974, 'reward_count_adherence_mean': 0.942307710647583, 'reward_count_adherence_std': 0.05439282953739166, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9885231256484985, 'reward_repeat_penalty_std': 0.012684613466262817, 'reward_near_duplicate_penalty_mean': 0.995418906211853, 'reward_near_duplicate_penalty_std': 0.0017438719514757395, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9930649995803833, 'reward_distinct_2_std': 0.01195761002600193, 'reward_judge_quality_mean': 0.17500001192092896, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.07325583696365356, 'reward_total_composite_std': 0.059081532061100006, 'epoch': 0.0}

  1%|          | 39/3300 [07:02<10:52:40, 12.01s/it]INFO 04-12 10:53:14 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 40/3300 [07:11<10:12:42, 11.28s/it]2026-04-12 10:53:23,731 | INFO | train_grpo_train | metrics_logged mode=train step=40

                                                    
{'loss': 0.0841, 'grad_norm': 14.31587028503418, 'learning_rate': 9.881818181818182e-06, 'num_tokens': 95077.0, 'completions/mean_length': 50.5, 'completions/min_length': 37.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.5, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.5635181665420532, 'rewards/meter/std': 0.396634966135025, '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/lexical_plausibility/mean': 0.9785957336425781, 'rewards/lexical_plausibility/std': 0.041514668613672256, 'rewards/judge_quality/mean': 0.3375000059604645, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.9940464496612549, 'rewards/repeat_penalty/std': 0.009807905182242393, 'rewards/near_duplicate_penalty/mean': 0.9940464496612549, 'rewards/near_duplicate_penalty/std': 0.009807905182242393, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1789971888065338, 'rewards/total_composite/std': 0.16574008762836456, 'reward': 0.1789971888065338, 'reward_std': 0.16574008762836456, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22694624960422516, 'sampling/sampling_logp_difference/max': 1.3110485076904297, 'sampling/importance_sampling_ratio/min': 0.2695372998714447, 'sampling/importance_sampling_ratio/mean': 1.0326288938522339, 'sampling/importance_sampling_ratio/max': 1.9913641214370728, 'entropy': 2.6005669236183167, 'clip_ratio/low_mean': 0.12105423025786877, 'clip_ratio/low_min': 0.12105423025786877, 'clip_ratio/high_mean': 0.08914576843380928, 'clip_ratio/high_max': 0.08914576843380928, 'clip_ratio/region_mean': 0.21019999869167805, 'reward_total_mean': 0.1789971888065338, 'reward_meter_mean': 0.5635181665420532, 'reward_meter_std': 0.396634966135025, '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_lexical_plausibility_mean': 0.9785957336425781, 'reward_lexical_plausibility_std': 0.041514668613672256, 'reward_repeat_penalty_mean': 0.9940464496612549, 'reward_repeat_penalty_std': 0.009807905182242393, 'reward_near_duplicate_penalty_mean': 0.9940464496612549, 'reward_near_duplicate_penalty_std': 0.009807905182242393, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3375000059604645, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.1789971888065338, 'reward_total_composite_std': 0.16574008762836456, 'epoch': 0.0}

  1%|          | 40/3300 [07:11<10:12:42, 11.28s/it]INFO 04-12 10:53:23 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 41/3300 [07:19<9:13:24, 10.19s/it] 2026-04-12 10:53:31,380 | INFO | train_grpo_train | metrics_logged mode=train step=41

                                                   
{'loss': -0.0119, 'grad_norm': 10.00015926361084, 'learning_rate': 9.87878787878788e-06, 'num_tokens': 96556.0, 'completions/mean_length': 28.875, 'completions/min_length': 28.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.875, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.7382493019104004, 'rewards/meter/std': 0.05320190265774727, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.9312499761581421, 'rewards/judge_quality/std': 0.015526460483670235, 'rewards/repeat_penalty/mean': 0.9941987991333008, 'rewards/repeat_penalty/std': 0.002123442478477955, 'rewards/near_duplicate_penalty/mean': 0.9941987991333008, 'rewards/near_duplicate_penalty/std': 0.002123442478477955, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.688033401966095, 'rewards/total_composite/std': 0.059163473546504974, 'reward': 0.688033401966095, 'reward_std': 0.059163469821214676, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04594084620475769, 'sampling/sampling_logp_difference/max': 1.053290605545044, 'sampling/importance_sampling_ratio/min': 0.34878814220428467, 'sampling/importance_sampling_ratio/mean': 1.0141326189041138, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2016809294000268, 'clip_ratio/low_mean': 0.013392857741564512, 'clip_ratio/low_min': 0.013392857741564512, 'clip_ratio/high_mean': 0.03191451169550419, 'clip_ratio/high_max': 0.03191451169550419, 'clip_ratio/region_mean': 0.0453073694370687, 'reward_total_mean': 0.688033401966095, 'reward_meter_mean': 0.7382493019104004, 'reward_meter_std': 0.05320190265774727, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9941987991333008, 'reward_repeat_penalty_std': 0.002123442478477955, 'reward_near_duplicate_penalty_mean': 0.9941987991333008, 'reward_near_duplicate_penalty_std': 0.002123442478477955, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.9312499761581421, 'reward_judge_quality_std': 0.015526460483670235, 'reward_total_composite_mean': 0.688033401966095, 'reward_total_composite_std': 0.059163473546504974, 'epoch': 0.0}

  1%|          | 41/3300 [07:19<9:13:24, 10.19s/it]INFO 04-12 10:53:31 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 42/3300 [07:34<10:33:27, 11.67s/it]2026-04-12 10:53:46,493 | INFO | train_grpo_train | metrics_logged mode=train step=42

                                                    
{'loss': 0.0701, 'grad_norm': 1.3855376243591309, 'learning_rate': 9.875757575757576e-06, 'num_tokens': 99223.0, 'completions/mean_length': 487.375, 'completions/min_length': 328.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.75, 'completions/mean_terminated_length': 413.5, 'completions/min_terminated_length': 328.0, 'completions/max_terminated_length': 499.0, 'rewards/meter/mean': 0.7093162536621094, 'rewards/meter/std': 0.1799566149711609, 'rewards/count_adherence/mean': 0.9666666984558105, 'rewards/count_adherence/std': 0.035634830594062805, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9994212985038757, 'rewards/lexical_plausibility/std': 0.0016368150245398283, 'rewards/judge_quality/mean': 0.21250000596046448, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.9619626998901367, 'rewards/repeat_penalty/std': 0.08446221798658371, 'rewards/near_duplicate_penalty/mean': 0.9942110180854797, 'rewards/near_duplicate_penalty/std': 0.004683594219386578, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9709505438804626, 'rewards/distinct_2/std': 0.0712062269449234, 'rewards/total_composite/mean': 0.10023243725299835, 'rewards/total_composite/std': 0.10760188847780228, 'reward': 0.10023243725299835, 'reward_std': 0.10760188102722168, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2188817411661148, 'sampling/sampling_logp_difference/max': 1.5691289901733398, 'sampling/importance_sampling_ratio/min': 0.20822647213935852, 'sampling/importance_sampling_ratio/mean': 1.0500988960266113, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7934285998344421, 'clip_ratio/low_mean': 0.017911585047841072, 'clip_ratio/low_min': 0.017911585047841072, 'clip_ratio/high_mean': 0.02655310556292534, 'clip_ratio/high_max': 0.02655310556292534, 'clip_ratio/region_mean': 0.04446469061076641, 'reward_total_mean': 0.10023243725299835, 'reward_meter_mean': 0.7093162536621094, 'reward_meter_std': 0.1799566149711609, 'reward_count_adherence_mean': 0.9666666984558105, 'reward_count_adherence_std': 0.035634830594062805, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9994212985038757, 'reward_lexical_plausibility_std': 0.0016368150245398283, 'reward_repeat_penalty_mean': 0.9619626998901367, 'reward_repeat_penalty_std': 0.08446221798658371, 'reward_near_duplicate_penalty_mean': 0.9942110180854797, 'reward_near_duplicate_penalty_std': 0.004683594219386578, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9709505438804626, 'reward_distinct_2_std': 0.0712062269449234, 'reward_judge_quality_mean': 0.21250000596046448, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.10023243725299835, 'reward_total_composite_std': 0.10760188847780228, 'epoch': 0.0}

  1%|▏         | 42/3300 [07:34<10:33:27, 11.67s/it]INFO 04-12 10:53:46 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 43/3300 [07:43<9:56:32, 10.99s/it] 2026-04-12 10:53:55,904 | INFO | train_grpo_train | metrics_logged mode=train step=43

                                                   
{'loss': 0.0607, 'grad_norm': 10.016197204589844, 'learning_rate': 9.872727272727274e-06, 'num_tokens': 101282.0, 'completions/mean_length': 87.375, 'completions/min_length': 66.0, 'completions/max_length': 107.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 87.375, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 107.0, 'rewards/meter/mean': 0.41793814301490784, 'rewards/meter/std': 0.41469404101371765, '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/lexical_plausibility/mean': 0.9853937029838562, 'rewards/lexical_plausibility/std': 0.02810603752732277, 'rewards/judge_quality/mean': 0.2875000238418579, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.9918968677520752, 'rewards/repeat_penalty/std': 0.008489455096423626, 'rewards/near_duplicate_penalty/mean': 0.9918968677520752, 'rewards/near_duplicate_penalty/std': 0.008489455096423626, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.13019166886806488, 'rewards/total_composite/std': 0.15118736028671265, 'reward': 0.13019166886806488, 'reward_std': 0.15118736028671265, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20895442366600037, 'sampling/sampling_logp_difference/max': 1.5569829940795898, 'sampling/importance_sampling_ratio/min': 0.21077100932598114, 'sampling/importance_sampling_ratio/mean': 1.0310027599334717, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.384279817342758, 'clip_ratio/low_mean': 0.09161348175257444, 'clip_ratio/low_min': 0.09161348175257444, 'clip_ratio/high_mean': 0.06697879917919636, 'clip_ratio/high_max': 0.06697879917919636, 'clip_ratio/region_mean': 0.1585922809317708, 'reward_total_mean': 0.13019166886806488, 'reward_meter_mean': 0.41793814301490784, 'reward_meter_std': 0.41469404101371765, '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_lexical_plausibility_mean': 0.9853937029838562, 'reward_lexical_plausibility_std': 0.02810603752732277, 'reward_repeat_penalty_mean': 0.9918968677520752, 'reward_repeat_penalty_std': 0.008489455096423626, 'reward_near_duplicate_penalty_mean': 0.9918968677520752, 'reward_near_duplicate_penalty_std': 0.008489455096423626, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.2875000238418579, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.13019166886806488, 'reward_total_composite_std': 0.15118736028671265, 'epoch': 0.0}

  1%|▏         | 43/3300 [07:43<9:56:32, 10.99s/it]INFO 04-12 10:53:56 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 44/3300 [07:53<9:41:57, 10.72s/it]2026-04-12 10:54:06,010 | INFO | train_grpo_train | metrics_logged mode=train step=44

                                                   
{'loss': 0.0253, 'grad_norm': 9.294432640075684, 'learning_rate': 9.869696969696971e-06, 'num_tokens': 103889.0, 'completions/mean_length': 150.875, 'completions/min_length': 103.0, 'completions/max_length': 197.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 150.875, 'completions/min_terminated_length': 103.0, 'completions/max_terminated_length': 197.0, 'rewards/meter/mean': 0.4813275933265686, 'rewards/meter/std': 0.4182776212692261, 'rewards/count_adherence/mean': 0.84375, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.08345229178667068, 'rewards/repeat_penalty/mean': 0.988006591796875, 'rewards/repeat_penalty/std': 0.015328135341405869, 'rewards/near_duplicate_penalty/mean': 0.9916410446166992, 'rewards/near_duplicate_penalty/std': 0.007457350846379995, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9963017702102661, 'rewards/distinct_2/std': 0.01046015229076147, 'rewards/total_composite/mean': 0.14278945326805115, 'rewards/total_composite/std': 0.12642918527126312, 'reward': 0.14278945326805115, 'reward_std': 0.12642918527126312, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21853402256965637, 'sampling/sampling_logp_difference/max': 3.2357258796691895, 'sampling/importance_sampling_ratio/min': 0.039331644773483276, 'sampling/importance_sampling_ratio/mean': 1.0230604410171509, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0845102220773697, 'clip_ratio/low_mean': 0.10745364427566528, 'clip_ratio/low_min': 0.10745364427566528, 'clip_ratio/high_mean': 0.07006077468395233, 'clip_ratio/high_max': 0.07006077468395233, 'clip_ratio/region_mean': 0.17751441895961761, 'reward_total_mean': 0.14278945326805115, 'reward_meter_mean': 0.4813275933265686, 'reward_meter_std': 0.4182776212692261, 'reward_count_adherence_mean': 0.84375, 'reward_count_adherence_std': 0.12938730418682098, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.988006591796875, 'reward_repeat_penalty_std': 0.015328135341405869, 'reward_near_duplicate_penalty_mean': 0.9916410446166992, 'reward_near_duplicate_penalty_std': 0.007457350846379995, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9963017702102661, 'reward_distinct_2_std': 0.01046015229076147, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.08345229178667068, 'reward_total_composite_mean': 0.14278945326805115, 'reward_total_composite_std': 0.12642918527126312, 'epoch': 0.0}

  1%|▏         | 44/3300 [07:53<9:41:57, 10.72s/it]INFO 04-12 10:54:06 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 45/3300 [08:03<9:28:46, 10.48s/it]2026-04-12 10:54:15,934 | INFO | train_grpo_train | metrics_logged mode=train step=45

                                                   
{'loss': 0.0626, 'grad_norm': 11.02723503112793, 'learning_rate': 9.866666666666668e-06, 'num_tokens': 105874.0, 'completions/mean_length': 75.125, 'completions/min_length': 51.0, 'completions/max_length': 106.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.125, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 106.0, 'rewards/meter/mean': 0.361456036567688, 'rewards/meter/std': 0.3149069845676422, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9866727590560913, 'rewards/lexical_plausibility/std': 0.02516147866845131, 'rewards/judge_quality/mean': 0.4375, 'rewards/judge_quality/std': 0.0353553369641304, 'rewards/repeat_penalty/mean': 0.9583264589309692, 'rewards/repeat_penalty/std': 0.04895475506782532, 'rewards/near_duplicate_penalty/mean': 0.9780596494674683, 'rewards/near_duplicate_penalty/std': 0.02558378502726555, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9793447256088257, 'rewards/distinct_2/std': 0.028561171144247055, 'rewards/total_composite/mean': 0.16231635212898254, 'rewards/total_composite/std': 0.14212195575237274, 'reward': 0.16231635212898254, 'reward_std': 0.14212194085121155, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19996686279773712, 'sampling/sampling_logp_difference/max': 1.33187997341156, 'sampling/importance_sampling_ratio/min': 0.26398053765296936, 'sampling/importance_sampling_ratio/mean': 1.028428554534912, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8198257088661194, 'clip_ratio/low_mean': 0.09802989661693573, 'clip_ratio/low_min': 0.09802989661693573, 'clip_ratio/high_mean': 0.09399807266891003, 'clip_ratio/high_max': 0.09399807266891003, 'clip_ratio/region_mean': 0.19202796928584576, 'reward_total_mean': 0.16231635212898254, 'reward_meter_mean': 0.361456036567688, 'reward_meter_std': 0.3149069845676422, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9866727590560913, 'reward_lexical_plausibility_std': 0.02516147866845131, 'reward_repeat_penalty_mean': 0.9583264589309692, 'reward_repeat_penalty_std': 0.04895475506782532, 'reward_near_duplicate_penalty_mean': 0.9780596494674683, 'reward_near_duplicate_penalty_std': 0.02558378502726555, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9793447256088257, 'reward_distinct_2_std': 0.028561171144247055, 'reward_judge_quality_mean': 0.4375, 'reward_judge_quality_std': 0.0353553369641304, 'reward_total_composite_mean': 0.16231635212898254, 'reward_total_composite_std': 0.14212195575237274, 'epoch': 0.0}

  1%|▏         | 45/3300 [08:03<9:28:46, 10.48s/it]INFO 04-12 10:54:16 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 46/3300 [08:12<8:58:24,  9.93s/it]2026-04-12 10:54:24,563 | INFO | train_grpo_train | metrics_logged mode=train step=46

                                                   
{'loss': -0.004, 'grad_norm': 16.038625717163086, 'learning_rate': 9.863636363636364e-06, 'num_tokens': 107225.0, 'completions/mean_length': 46.875, 'completions/min_length': 28.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.875, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.4543209373950958, 'rewards/meter/std': 0.4082188904285431, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9658395051956177, 'rewards/lexical_plausibility/std': 0.06418022513389587, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.1414213478565216, 'rewards/repeat_penalty/mean': 0.9392427206039429, 'rewards/repeat_penalty/std': 0.11786352097988129, 'rewards/near_duplicate_penalty/mean': 0.9392427206039429, 'rewards/near_duplicate_penalty/std': 0.11786352097988129, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.11724015325307846, 'rewards/total_composite/std': 0.14281949400901794, 'reward': 0.11724015325307846, 'reward_std': 0.14281950891017914, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.262806236743927, 'sampling/sampling_logp_difference/max': 1.9793095588684082, 'sampling/importance_sampling_ratio/min': 0.13816459476947784, 'sampling/importance_sampling_ratio/mean': 1.0275317430496216, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.951196700334549, 'clip_ratio/low_mean': 0.1215451406314969, 'clip_ratio/low_min': 0.1215451406314969, 'clip_ratio/high_mean': 0.1346866339445114, 'clip_ratio/high_max': 0.1346866339445114, 'clip_ratio/region_mean': 0.2562317745760083, 'reward_total_mean': 0.11724015325307846, 'reward_meter_mean': 0.4543209373950958, 'reward_meter_std': 0.4082188904285431, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9658395051956177, 'reward_lexical_plausibility_std': 0.06418022513389587, 'reward_repeat_penalty_mean': 0.9392427206039429, 'reward_repeat_penalty_std': 0.11786352097988129, 'reward_near_duplicate_penalty_mean': 0.9392427206039429, 'reward_near_duplicate_penalty_std': 0.11786352097988129, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.1414213478565216, 'reward_total_composite_mean': 0.11724015325307846, 'reward_total_composite_std': 0.14281949400901794, 'epoch': 0.0}

  1%|▏         | 46/3300 [08:12<8:58:24,  9.93s/it]INFO 04-12 10:54:24 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 47/3300 [08:22<8:51:57,  9.81s/it]2026-04-12 10:54:34,105 | INFO | train_grpo_train | metrics_logged mode=train step=47

                                                   
{'loss': 0.112, 'grad_norm': 23.37023162841797, 'learning_rate': 9.860606060606061e-06, 'num_tokens': 108866.0, 'completions/mean_length': 35.125, 'completions/min_length': 30.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.125, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.5010475516319275, 'rewards/meter/std': 0.38557562232017517, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.39625000953674316, 'rewards/judge_quality/std': 0.24136146903038025, 'rewards/repeat_penalty/mean': 0.9946637153625488, 'rewards/repeat_penalty/std': 0.007449817378073931, 'rewards/near_duplicate_penalty/mean': 0.9946637153625488, 'rewards/near_duplicate_penalty/std': 0.007449817378073931, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2251858413219452, 'rewards/total_composite/std': 0.29974380135536194, 'reward': 0.2251858413219452, 'reward_std': 0.29974377155303955, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.26643556356430054, 'sampling/sampling_logp_difference/max': 1.6169767379760742, 'sampling/importance_sampling_ratio/min': 0.19849790632724762, 'sampling/importance_sampling_ratio/mean': 1.0140801668167114, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5216663479804993, 'clip_ratio/low_mean': 0.13808361627161503, 'clip_ratio/low_min': 0.13808361627161503, 'clip_ratio/high_mean': 0.0861111138947308, 'clip_ratio/high_max': 0.0861111138947308, 'clip_ratio/region_mean': 0.22419473016634583, 'reward_total_mean': 0.2251858413219452, 'reward_meter_mean': 0.5010475516319275, 'reward_meter_std': 0.38557562232017517, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9946637153625488, 'reward_repeat_penalty_std': 0.007449817378073931, 'reward_near_duplicate_penalty_mean': 0.9946637153625488, 'reward_near_duplicate_penalty_std': 0.007449817378073931, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.39625000953674316, 'reward_judge_quality_std': 0.24136146903038025, 'reward_total_composite_mean': 0.2251858413219452, 'reward_total_composite_std': 0.29974380135536194, 'epoch': 0.0}

  1%|▏         | 47/3300 [08:22<8:51:57,  9.81s/it]INFO 04-12 10:54:34 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 48/3300 [08:31<8:47:44,  9.74s/it]2026-04-12 10:54:43,667 | INFO | train_grpo_train | metrics_logged mode=train step=48

                                                   
{'loss': 0.0071, 'grad_norm': 9.893824577331543, 'learning_rate': 9.857575757575758e-06, 'num_tokens': 111117.0, 'completions/mean_length': 104.375, 'completions/min_length': 82.0, 'completions/max_length': 150.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 104.375, 'completions/min_terminated_length': 82.0, 'completions/max_terminated_length': 150.0, 'rewards/meter/mean': 0.806372344493866, 'rewards/meter/std': 0.2834595739841461, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.06408698856830597, 'rewards/repeat_penalty/mean': 0.9797012209892273, 'rewards/repeat_penalty/std': 0.019535429775714874, 'rewards/near_duplicate_penalty/mean': 0.9892376661300659, 'rewards/near_duplicate_penalty/std': 0.008412816561758518, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9903604984283447, 'rewards/distinct_2/std': 0.017878584563732147, 'rewards/total_composite/mean': 0.2767994999885559, 'rewards/total_composite/std': 0.11022676527500153, 'reward': 0.2767994999885559, 'reward_std': 0.11022675782442093, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17351074516773224, 'sampling/sampling_logp_difference/max': 1.263441562652588, 'sampling/importance_sampling_ratio/min': 0.2826794683933258, 'sampling/importance_sampling_ratio/mean': 1.019784927368164, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6085834950208664, 'clip_ratio/low_mean': 0.07410561852157116, 'clip_ratio/low_min': 0.07410561852157116, 'clip_ratio/high_mean': 0.09488278347998857, 'clip_ratio/high_max': 0.09488278347998857, 'clip_ratio/region_mean': 0.16898840200155973, 'reward_total_mean': 0.2767994999885559, 'reward_meter_mean': 0.806372344493866, 'reward_meter_std': 0.2834595739841461, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9797012209892273, 'reward_repeat_penalty_std': 0.019535429775714874, 'reward_near_duplicate_penalty_mean': 0.9892376661300659, 'reward_near_duplicate_penalty_std': 0.008412816561758518, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9903604984283447, 'reward_distinct_2_std': 0.017878584563732147, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.06408698856830597, 'reward_total_composite_mean': 0.2767994999885559, 'reward_total_composite_std': 0.11022676527500153, 'epoch': 0.0}

  1%|▏         | 48/3300 [08:31<8:47:44,  9.74s/it]INFO 04-12 10:54:43 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 49/3300 [08:41<8:47:03,  9.73s/it]2026-04-12 10:54:53,373 | INFO | train_grpo_train | metrics_logged mode=train step=49

                                                   
{'loss': 0.0709, 'grad_norm': 14.83174991607666, 'learning_rate': 9.854545454545456e-06, 'num_tokens': 112936.0, 'completions/mean_length': 63.375, 'completions/min_length': 45.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.375, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.3535194396972656, 'rewards/meter/std': 0.29302817583084106, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.11649646610021591, 'rewards/repeat_penalty/mean': 0.9892464876174927, 'rewards/repeat_penalty/std': 0.006524175871163607, 'rewards/near_duplicate_penalty/mean': 0.9892464876174927, 'rewards/near_duplicate_penalty/std': 0.006524175871163607, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1120145320892334, 'rewards/total_composite/std': 0.09365059435367584, 'reward': 0.1120145320892334, 'reward_std': 0.09365059435367584, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2412882298231125, 'sampling/sampling_logp_difference/max': 2.8684377670288086, 'sampling/importance_sampling_ratio/min': 0.05678757280111313, 'sampling/importance_sampling_ratio/mean': 1.0413163900375366, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.34032841026783, 'clip_ratio/low_mean': 0.12963778898119926, 'clip_ratio/low_min': 0.12963778898119926, 'clip_ratio/high_mean': 0.09714464657008648, 'clip_ratio/high_max': 0.09714464657008648, 'clip_ratio/region_mean': 0.22678243555128574, 'reward_total_mean': 0.1120145320892334, 'reward_meter_mean': 0.3535194396972656, 'reward_meter_std': 0.29302817583084106, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9892464876174927, 'reward_repeat_penalty_std': 0.006524175871163607, 'reward_near_duplicate_penalty_mean': 0.9892464876174927, 'reward_near_duplicate_penalty_std': 0.006524175871163607, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.11649646610021591, 'reward_total_composite_mean': 0.1120145320892334, 'reward_total_composite_std': 0.09365059435367584, 'epoch': 0.0}

  1%|▏         | 49/3300 [08:41<8:47:03,  9.73s/it]INFO 04-12 10:54:53 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 50/3300 [08:51<8:54:09,  9.86s/it]2026-04-12 10:55:03,565 | INFO | train_grpo_train | metrics_logged mode=train step=50

                                                   
{'loss': 0.0261, 'grad_norm': 17.06206512451172, 'learning_rate': 9.851515151515151e-06, 'num_tokens': 114662.0, 'completions/mean_length': 58.75, 'completions/min_length': 48.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.75, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.3583456873893738, 'rewards/meter/std': 0.37512892484664917, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.48749998211860657, 'rewards/judge_quality/std': 0.14078859984874725, 'rewards/repeat_penalty/mean': 0.9947724342346191, 'rewards/repeat_penalty/std': 0.006623376626521349, 'rewards/near_duplicate_penalty/mean': 0.9947724342346191, 'rewards/near_duplicate_penalty/std': 0.006623376626521349, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.18473929166793823, 'rewards/total_composite/std': 0.21375440061092377, 'reward': 0.18473929166793823, 'reward_std': 0.21375440061092377, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21507924795150757, 'sampling/sampling_logp_difference/max': 1.6833810806274414, 'sampling/importance_sampling_ratio/min': 0.18574489653110504, 'sampling/importance_sampling_ratio/mean': 1.0162782669067383, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.93431057035923, 'clip_ratio/low_mean': 0.12212134897708893, 'clip_ratio/low_min': 0.12212134897708893, 'clip_ratio/high_mean': 0.03997507132589817, 'clip_ratio/high_max': 0.03997507132589817, 'clip_ratio/region_mean': 0.1620964203029871, 'reward_total_mean': 0.18473929166793823, 'reward_meter_mean': 0.3583456873893738, 'reward_meter_std': 0.37512892484664917, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.9947724342346191, 'reward_repeat_penalty_std': 0.006623376626521349, 'reward_near_duplicate_penalty_mean': 0.9947724342346191, 'reward_near_duplicate_penalty_std': 0.006623376626521349, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.48749998211860657, 'reward_judge_quality_std': 0.14078859984874725, 'reward_total_composite_mean': 0.18473929166793823, 'reward_total_composite_std': 0.21375440061092377, 'epoch': 0.0}

  2%|▏         | 50/3300 [08:51<8:54:09,  9.86s/it]INFO 04-12 10:55:03 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


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

 15%|█▌        | 2/13 [00:09<00:54,  4.98s/it]

 23%|██▎       | 3/13 [00:18<01:04,  6.50s/it]

 31%|███       | 4/13 [00:31<01:19,  8.78s/it]

 38%|███▊      | 5/13 [00:40<01:11,  8.99s/it]

 46%|████▌     | 6/13 [00:54<01:13, 10.56s/it]

 54%|█████▍    | 7/13 [01:03<01:00, 10.03s/it]

 62%|██████▏   | 8/13 [01:15<00:53, 10.66s/it]

 69%|██████▉   | 9/13 [01:28<00:45, 11.31s/it]

 77%|███████▋  | 10/13 [01:37<00:32, 10.78s/it]

 85%|████████▍ | 11/13 [01:48<00:21, 10.94s/it]

 92%|█████████▏| 12/13 [01:59<00:10, 10.98s/it]

100%|██████████| 13/13 [02:12<00:00, 11.36s/it]2026-04-12 10:57:26,405 | INFO | train_grpo_train | metrics_logged mode=eval step=50

                                                   

                                               
{'eval_loss': nan, 'eval_runtime': 142.8323, 'eval_samples_per_second': 0.728, 'eval_steps_per_second': 0.091, 'eval_num_tokens': 114662.0, 'eval_completions/mean_length': 180.56730769230768, 'eval_completions/min_length': 37.15384615384615, 'eval_completions/max_length': 402.84615384615387, 'eval_completions/clipped_ratio': 0.057692307692307696, 'eval_completions/mean_terminated_length': 159.45604647122897, 'eval_completions/min_terminated_length': 37.15384615384615, 'eval_completions/max_terminated_length': 339.61538461538464, 'eval_rewards/meter/mean': 0.37956270346274745, 'eval_rewards/meter/std': 0.3265574127435684, 'eval_rewards/count_adherence/mean': 0.9397292916591351, 'eval_rewards/count_adherence/std': 0.08025195621527158, 'eval_rewards/arabic_clean/mean': 0.9326923076923077, 'eval_rewards/arabic_clean/std': 0.13402181176038888, 'eval_rewards/lexical_plausibility/mean': 0.9808496099251968, 'eval_rewards/lexical_plausibility/std': 0.04648121876212267, 'eval_rewards/judge_quality/mean': 0.3682692234332745, 'eval_rewards/judge_quality/std': 0.18312470041788542, 'eval_rewards/repeat_penalty/mean': 0.9772035846343408, 'eval_rewards/repeat_penalty/std': 0.03686572052538395, 'eval_rewards/near_duplicate_penalty/mean': 0.9911981270863459, 'eval_rewards/near_duplicate_penalty/std': 0.011277515387449127, 'eval_rewards/opening_diversity/mean': 0.9927884615384616, 'eval_rewards/opening_diversity/std': 0.020397310646680687, 'eval_rewards/distinct_2/mean': 0.9927693880521334, 'eval_rewards/distinct_2/std': 0.014553762112672512, 'eval_rewards/total_composite/mean': 0.11893160469256915, 'eval_rewards/total_composite/std': 0.13336633231777412, 'eval_reward': 0.11893160469256915, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.14144953741477087, 'eval_sampling/sampling_logp_difference/max': 1.1957824413593, 'eval_sampling/importance_sampling_ratio/min': 0.3068786939749351, 'eval_sampling/importance_sampling_ratio/mean': 1.039452286866995, 'eval_sampling/importance_sampling_ratio/max': 1.6445801258087158, 'eval_entropy': 2.337759989958543, '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': 0.11893160469256915, 'eval_reward_meter_mean': 0.37956270346274745, 'eval_reward_meter_std': 0.3265574127435684, 'eval_reward_count_adherence_mean': 0.9397292916591351, 'eval_reward_count_adherence_std': 0.08025195621527158, 'eval_reward_arabic_clean_mean': 0.9326923076923077, 'eval_reward_arabic_clean_std': 0.13402181176038888, 'eval_reward_lexical_plausibility_mean': 0.9808496099251968, 'eval_reward_lexical_plausibility_std': 0.04648121876212267, 'eval_reward_repeat_penalty_mean': 0.9772035846343408, 'eval_reward_repeat_penalty_std': 0.03686572052538395, 'eval_reward_near_duplicate_penalty_mean': 0.9911981270863459, 'eval_reward_near_duplicate_penalty_std': 0.011277515387449127, 'eval_reward_opening_diversity_mean': 0.9927884615384616, 'eval_reward_opening_diversity_std': 0.020397310646680687, 'eval_reward_distinct_2_mean': 0.9927693880521334, 'eval_reward_distinct_2_std': 0.014553762112672512, 'eval_reward_judge_quality_mean': 0.3682692234332745, 'eval_reward_judge_quality_std': 0.18312470041788542, 'eval_reward_total_composite_mean': 0.11893160469256915, 'eval_reward_total_composite_std': 0.13336633231777412, 'epoch': 0.0}

  2%|▏         | 50/3300 [11:14<8:54:09,  9.86s/it]

100%|██████████| 13/13 [02:12<00:00, 11.36s/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(

  2%|▏         | 51/3300 [11:27<48:35:18, 53.84s/it]2026-04-12 10:57:39,996 | INFO | train_grpo_train | metrics_logged mode=train step=51

                                                    
{'loss': -0.0337, 'grad_norm': 10.903347969055176, 'learning_rate': 9.84848484848485e-06, 'num_tokens': 116950.0, 'completions/mean_length': 113.0, 'completions/min_length': 69.0, 'completions/max_length': 170.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 113.0, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 170.0, 'rewards/meter/mean': 0.28861385583877563, 'rewards/meter/std': 0.36539360880851746, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.9817765951156616, 'rewards/repeat_penalty/std': 0.01925649307668209, 'rewards/near_duplicate_penalty/mean': 0.9861289858818054, 'rewards/near_duplicate_penalty/std': 0.009517847560346127, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9955277442932129, 'rewards/distinct_2/std': 0.012649492360651493, 'rewards/total_composite/mean': 0.09108218550682068, 'rewards/total_composite/std': 0.108382448554039, 'reward': 0.09108218550682068, 'reward_std': 0.1083824560046196, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2257484644651413, 'sampling/sampling_logp_difference/max': 1.6658835411071777, 'sampling/importance_sampling_ratio/min': 0.18902356922626495, 'sampling/importance_sampling_ratio/mean': 1.0398006439208984, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.419378399848938, 'clip_ratio/low_mean': 0.1551432702690363, 'clip_ratio/low_min': 0.1551432702690363, 'clip_ratio/high_mean': 0.04990053176879883, 'clip_ratio/high_max': 0.04990053176879883, 'clip_ratio/region_mean': 0.20504380203783512, 'reward_total_mean': 0.09108218550682068, 'reward_meter_mean': 0.28861385583877563, 'reward_meter_std': 0.36539360880851746, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9817765951156616, 'reward_repeat_penalty_std': 0.01925649307668209, 'reward_near_duplicate_penalty_mean': 0.9861289858818054, 'reward_near_duplicate_penalty_std': 0.009517847560346127, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9955277442932129, 'reward_distinct_2_std': 0.012649492360651493, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.09108218550682068, 'reward_total_composite_std': 0.108382448554039, 'epoch': 0.0}

  2%|▏         | 51/3300 [11:27<48:35:18, 53.84s/it]INFO 04-12 10:57:40 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 52/3300 [11:36<36:15:35, 40.19s/it]2026-04-12 10:57:48,340 | INFO | train_grpo_train | metrics_logged mode=train step=52

                                                    
{'loss': -0.0877, 'grad_norm': 24.892927169799805, 'learning_rate': 9.845454545454546e-06, 'num_tokens': 118357.0, 'completions/mean_length': 22.875, 'completions/min_length': 15.0, 'completions/max_length': 31.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.875, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.5366003513336182, 'rewards/meter/std': 0.40989452600479126, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.6974999904632568, 'rewards/judge_quality/std': 0.3259163200855255, 'rewards/repeat_penalty/mean': 0.7438346147537231, 'rewards/repeat_penalty/std': 0.011416047811508179, 'rewards/near_duplicate_penalty/mean': 0.9917795062065125, 'rewards/near_duplicate_penalty/std': 0.015221425332129002, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.43747198581695557, 'rewards/total_composite/std': 0.36508339643478394, 'reward': 0.43747198581695557, 'reward_std': 0.36508339643478394, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21967212855815887, 'sampling/sampling_logp_difference/max': 1.12404203414917, 'sampling/importance_sampling_ratio/min': 0.32496359944343567, 'sampling/importance_sampling_ratio/mean': 1.024284839630127, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9608293920755386, 'clip_ratio/low_mean': 0.1209017550572753, 'clip_ratio/low_min': 0.1209017550572753, 'clip_ratio/high_mean': 0.08847926370799541, 'clip_ratio/high_max': 0.08847926370799541, 'clip_ratio/region_mean': 0.2093810187652707, 'reward_total_mean': 0.43747198581695557, 'reward_meter_mean': 0.5366003513336182, 'reward_meter_std': 0.40989452600479126, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.7438346147537231, 'reward_repeat_penalty_std': 0.011416047811508179, 'reward_near_duplicate_penalty_mean': 0.9917795062065125, 'reward_near_duplicate_penalty_std': 0.015221425332129002, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6974999904632568, 'reward_judge_quality_std': 0.3259163200855255, 'reward_total_composite_mean': 0.43747198581695557, 'reward_total_composite_std': 0.36508339643478394, 'epoch': 0.0}

  2%|▏         | 52/3300 [11:36<36:15:35, 40.19s/it]INFO 04-12 10:57:48 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 53/3300 [11:46<28:13:46, 31.30s/it]2026-04-12 10:57:58,893 | INFO | train_grpo_train | metrics_logged mode=train step=53

                                                    
{'loss': 0.085, 'grad_norm': 8.417654991149902, 'learning_rate': 9.842424242424243e-06, 'num_tokens': 121027.0, 'completions/mean_length': 154.75, 'completions/min_length': 120.0, 'completions/max_length': 189.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 154.75, 'completions/min_terminated_length': 120.0, 'completions/max_terminated_length': 189.0, 'rewards/meter/mean': 0.45218032598495483, 'rewards/meter/std': 0.35878920555114746, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.1035098284482956, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9745864868164062, 'rewards/lexical_plausibility/std': 0.05568007379770279, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.9912744760513306, 'rewards/repeat_penalty/std': 0.011122609488666058, 'rewards/near_duplicate_penalty/mean': 0.9947509765625, 'rewards/near_duplicate_penalty/std': 0.004814809188246727, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.996503472328186, 'rewards/distinct_2/std': 0.00988961011171341, 'rewards/total_composite/mean': 0.09445170313119888, 'rewards/total_composite/std': 0.08176426589488983, 'reward': 0.09445170313119888, 'reward_std': 0.08176426589488983, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22425159811973572, 'sampling/sampling_logp_difference/max': 1.8631505966186523, 'sampling/importance_sampling_ratio/min': 0.15518295764923096, 'sampling/importance_sampling_ratio/mean': 1.0403592586517334, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.76952862739563, 'clip_ratio/low_mean': 0.10465529002249241, 'clip_ratio/low_min': 0.10465529002249241, 'clip_ratio/high_mean': 0.08321425691246986, 'clip_ratio/high_max': 0.08321425691246986, 'clip_ratio/region_mean': 0.18786954693496227, 'reward_total_mean': 0.09445170313119888, 'reward_meter_mean': 0.45218032598495483, 'reward_meter_std': 0.35878920555114746, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.1035098284482956, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9745864868164062, 'reward_lexical_plausibility_std': 0.05568007379770279, 'reward_repeat_penalty_mean': 0.9912744760513306, 'reward_repeat_penalty_std': 0.011122609488666058, 'reward_near_duplicate_penalty_mean': 0.9947509765625, 'reward_near_duplicate_penalty_std': 0.004814809188246727, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.996503472328186, 'reward_distinct_2_std': 0.00988961011171341, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.09445170313119888, 'reward_total_composite_std': 0.08176426589488983, 'epoch': 0.0}

  2%|▏         | 53/3300 [11:46<28:13:46, 31.30s/it]INFO 04-12 10:57:59 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 54/3300 [11:56<22:26:51, 24.90s/it]2026-04-12 10:58:08,851 | INFO | train_grpo_train | metrics_logged mode=train step=54

                                                    
{'loss': 0.0818, 'grad_norm': 12.048750877380371, 'learning_rate': 9.83939393939394e-06, 'num_tokens': 122732.0, 'completions/mean_length': 49.125, 'completions/min_length': 36.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.125, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.5787004232406616, 'rewards/meter/std': 0.38724616169929504, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.45875000953674316, 'rewards/judge_quality/std': 0.192831352353096, 'rewards/repeat_penalty/mean': 0.9991129636764526, 'rewards/repeat_penalty/std': 0.0021831849589943886, 'rewards/near_duplicate_penalty/mean': 0.9991129636764526, 'rewards/near_duplicate_penalty/std': 0.0021831849589943886, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2784807085990906, 'rewards/total_composite/std': 0.27733859419822693, 'reward': 0.2784807085990906, 'reward_std': 0.27733856439590454, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22375231981277466, 'sampling/sampling_logp_difference/max': 1.4279108047485352, 'sampling/importance_sampling_ratio/min': 0.23980940878391266, 'sampling/importance_sampling_ratio/mean': 1.0389413833618164, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.739379346370697, 'clip_ratio/low_mean': 0.12489048577845097, 'clip_ratio/low_min': 0.12489048577845097, 'clip_ratio/high_mean': 0.06473684310913086, 'clip_ratio/high_max': 0.06473684310913086, 'clip_ratio/region_mean': 0.18962732888758183, 'reward_total_mean': 0.2784807085990906, 'reward_meter_mean': 0.5787004232406616, 'reward_meter_std': 0.38724616169929504, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9991129636764526, 'reward_repeat_penalty_std': 0.0021831849589943886, 'reward_near_duplicate_penalty_mean': 0.9991129636764526, 'reward_near_duplicate_penalty_std': 0.0021831849589943886, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.45875000953674316, 'reward_judge_quality_std': 0.192831352353096, 'reward_total_composite_mean': 0.2784807085990906, 'reward_total_composite_std': 0.27733859419822693, 'epoch': 0.0}

  2%|▏         | 54/3300 [11:56<22:26:51, 24.90s/it]INFO 04-12 10:58:09 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 55/3300 [12:12<19:50:38, 22.01s/it]2026-04-12 10:58:24,141 | INFO | train_grpo_train | metrics_logged mode=train step=55

                                                    
{'loss': 0.0738, 'grad_norm': 2.989703416824341, 'learning_rate': 9.836363636363637e-06, 'num_tokens': 126060.0, 'completions/mean_length': 463.0, 'completions/min_length': 304.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 414.0, 'completions/min_terminated_length': 304.0, 'completions/max_terminated_length': 492.0, 'rewards/meter/mean': 0.4480252265930176, 'rewards/meter/std': 0.19926531612873077, 'rewards/count_adherence/mean': 0.953125, 'rewards/count_adherence/std': 0.04419417306780815, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9838997721672058, 'rewards/lexical_plausibility/std': 0.0295176412910223, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.9866570234298706, 'rewards/repeat_penalty/std': 0.011641118675470352, 'rewards/near_duplicate_penalty/mean': 0.9968233108520508, 'rewards/near_duplicate_penalty/std': 0.002474944805726409, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9897903203964233, 'rewards/distinct_2/std': 0.010273823514580727, 'rewards/total_composite/mean': 0.05033401772379875, 'rewards/total_composite/std': 0.04285261407494545, 'reward': 0.05033401772379875, 'reward_std': 0.04285261407494545, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23841920495033264, 'sampling/sampling_logp_difference/max': 1.7952909469604492, 'sampling/importance_sampling_ratio/min': 0.16607911884784698, 'sampling/importance_sampling_ratio/mean': 1.0540883541107178, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.888288974761963, 'clip_ratio/low_mean': 0.06659147888422012, 'clip_ratio/low_min': 0.06659147888422012, 'clip_ratio/high_mean': 0.02017114870250225, 'clip_ratio/high_max': 0.02017114870250225, 'clip_ratio/region_mean': 0.08676262758672237, 'reward_total_mean': 0.05033401772379875, 'reward_meter_mean': 0.4480252265930176, 'reward_meter_std': 0.19926531612873077, 'reward_count_adherence_mean': 0.953125, 'reward_count_adherence_std': 0.04419417306780815, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9838997721672058, 'reward_lexical_plausibility_std': 0.0295176412910223, 'reward_repeat_penalty_mean': 0.9866570234298706, 'reward_repeat_penalty_std': 0.011641118675470352, 'reward_near_duplicate_penalty_mean': 0.9968233108520508, 'reward_near_duplicate_penalty_std': 0.002474944805726409, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9897903203964233, 'reward_distinct_2_std': 0.010273823514580727, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.05033401772379875, 'reward_total_composite_std': 0.04285261407494545, 'epoch': 0.0}

  2%|▏         | 55/3300 [12:12<19:50:38, 22.01s/it]INFO 04-12 10:58:24 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 56/3300 [12:21<16:29:31, 18.30s/it]2026-04-12 10:58:33,780 | INFO | train_grpo_train | metrics_logged mode=train step=56

                                                    
{'loss': 0.0507, 'grad_norm': 17.793067932128906, 'learning_rate': 9.833333333333333e-06, 'num_tokens': 127700.0, 'completions/mean_length': 45.0, 'completions/min_length': 34.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.0, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.5145763754844666, 'rewards/meter/std': 0.37867268919944763, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9702739715576172, 'rewards/lexical_plausibility/std': 0.06984595954418182, 'rewards/judge_quality/mean': 0.3837500214576721, 'rewards/judge_quality/std': 0.24663087725639343, 'rewards/repeat_penalty/mean': 0.987379789352417, 'rewards/repeat_penalty/std': 0.018242988735437393, 'rewards/near_duplicate_penalty/mean': 0.987379789352417, 'rewards/near_duplicate_penalty/std': 0.018242988735437393, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2455739825963974, 'rewards/total_composite/std': 0.2809582054615021, 'reward': 0.2455739825963974, 'reward_std': 0.28095823526382446, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22564570605754852, 'sampling/sampling_logp_difference/max': 1.7728815078735352, 'sampling/importance_sampling_ratio/min': 0.16984286904335022, 'sampling/importance_sampling_ratio/mean': 1.0182222127914429, 'sampling/importance_sampling_ratio/max': 1.9703538417816162, 'entropy': 2.5086857080459595, 'clip_ratio/low_mean': 0.11486524250358343, 'clip_ratio/low_min': 0.11486524250358343, 'clip_ratio/high_mean': 0.1189475804567337, 'clip_ratio/high_max': 0.1189475804567337, 'clip_ratio/region_mean': 0.23381282296031713, 'reward_total_mean': 0.2455739825963974, 'reward_meter_mean': 0.5145763754844666, 'reward_meter_std': 0.37867268919944763, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9702739715576172, 'reward_lexical_plausibility_std': 0.06984595954418182, 'reward_repeat_penalty_mean': 0.987379789352417, 'reward_repeat_penalty_std': 0.018242988735437393, 'reward_near_duplicate_penalty_mean': 0.987379789352417, 'reward_near_duplicate_penalty_std': 0.018242988735437393, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3837500214576721, 'reward_judge_quality_std': 0.24663087725639343, 'reward_total_composite_mean': 0.2455739825963974, 'reward_total_composite_std': 0.2809582054615021, 'epoch': 0.0}

  2%|▏         | 56/3300 [12:21<16:29:31, 18.30s/it]INFO 04-12 10:58:34 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 57/3300 [12:31<14:17:11, 15.86s/it]2026-04-12 10:58:43,941 | INFO | train_grpo_train | metrics_logged mode=train step=57

                                                    
{'loss': 0.0935, 'grad_norm': 13.650352478027344, 'learning_rate': 9.830303030303032e-06, 'num_tokens': 129539.0, 'completions/mean_length': 58.875, 'completions/min_length': 40.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.875, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.8692176938056946, 'rewards/meter/std': 0.30586138367652893, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9266800880432129, 'rewards/lexical_plausibility/std': 0.12982980906963348, 'rewards/judge_quality/mean': 0.5174999833106995, 'rewards/judge_quality/std': 0.26868730783462524, 'rewards/repeat_penalty/mean': 0.9955106973648071, 'rewards/repeat_penalty/std': 0.00844684150069952, 'rewards/near_duplicate_penalty/mean': 0.9955106973648071, 'rewards/near_duplicate_penalty/std': 0.00844684150069952, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.45738109946250916, 'rewards/total_composite/std': 0.3089926540851593, 'reward': 0.45738109946250916, 'reward_std': 0.3089926540851593, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21686285734176636, 'sampling/sampling_logp_difference/max': 1.2738847732543945, 'sampling/importance_sampling_ratio/min': 0.2797427475452423, 'sampling/importance_sampling_ratio/mean': 1.0531916618347168, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.3706904649734497, 'clip_ratio/low_mean': 0.1262198593467474, 'clip_ratio/low_min': 0.1262198593467474, 'clip_ratio/high_mean': 0.049512987956404686, 'clip_ratio/high_max': 0.049512987956404686, 'clip_ratio/region_mean': 0.17573284730315208, 'reward_total_mean': 0.45738109946250916, 'reward_meter_mean': 0.8692176938056946, 'reward_meter_std': 0.30586138367652893, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9266800880432129, 'reward_lexical_plausibility_std': 0.12982980906963348, 'reward_repeat_penalty_mean': 0.9955106973648071, 'reward_repeat_penalty_std': 0.00844684150069952, 'reward_near_duplicate_penalty_mean': 0.9955106973648071, 'reward_near_duplicate_penalty_std': 0.00844684150069952, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5174999833106995, 'reward_judge_quality_std': 0.26868730783462524, 'reward_total_composite_mean': 0.45738109946250916, 'reward_total_composite_std': 0.3089926540851593, 'epoch': 0.0}

  2%|▏         | 57/3300 [12:31<14:17:11, 15.86s/it]INFO 04-12 10:58:44 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 58/3300 [12:41<12:32:46, 13.93s/it]2026-04-12 10:58:53,375 | INFO | train_grpo_train | metrics_logged mode=train step=58

                                                    
{'loss': 0.039, 'grad_norm': 17.057424545288086, 'learning_rate': 9.827272727272729e-06, 'num_tokens': 131234.0, 'completions/mean_length': 44.875, 'completions/min_length': 33.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.875, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.4741390645503998, 'rewards/meter/std': 0.38624247908592224, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.48000001907348633, 'rewards/judge_quality/std': 0.29871153831481934, 'rewards/repeat_penalty/mean': 0.9705674648284912, 'rewards/repeat_penalty/std': 0.040225934237241745, 'rewards/near_duplicate_penalty/mean': 0.9705674648284912, 'rewards/near_duplicate_penalty/std': 0.040225934237241745, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2696312367916107, 'rewards/total_composite/std': 0.37265121936798096, 'reward': 0.2696312367916107, 'reward_std': 0.37265118956565857, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23961296677589417, 'sampling/sampling_logp_difference/max': 1.6894340515136719, 'sampling/importance_sampling_ratio/min': 0.18462397158145905, 'sampling/importance_sampling_ratio/mean': 1.0087285041809082, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.136370226740837, 'clip_ratio/low_mean': 0.11129444930702448, 'clip_ratio/low_min': 0.11129444930702448, 'clip_ratio/high_mean': 0.04762791842222214, 'clip_ratio/high_max': 0.04762791842222214, 'clip_ratio/region_mean': 0.15892236772924662, 'reward_total_mean': 0.2696312367916107, 'reward_meter_mean': 0.4741390645503998, 'reward_meter_std': 0.38624247908592224, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9705674648284912, 'reward_repeat_penalty_std': 0.040225934237241745, 'reward_near_duplicate_penalty_mean': 0.9705674648284912, 'reward_near_duplicate_penalty_std': 0.040225934237241745, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.48000001907348633, 'reward_judge_quality_std': 0.29871153831481934, 'reward_total_composite_mean': 0.2696312367916107, 'reward_total_composite_std': 0.37265121936798096, 'epoch': 0.0}

  2%|▏         | 58/3300 [12:41<12:32:46, 13.93s/it]INFO 04-12 10:58:53 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 59/3300 [12:50<11:07:25, 12.36s/it]2026-04-12 10:59:02,054 | INFO | train_grpo_train | metrics_logged mode=train step=59

                                                    
{'loss': 0.0104, 'grad_norm': 17.808332443237305, 'learning_rate': 9.824242424242425e-06, 'num_tokens': 132754.0, 'completions/mean_length': 35.0, 'completions/min_length': 33.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.0, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.8920453786849976, 'rewards/meter/std': 0.04058147594332695, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.8725000023841858, 'rewards/judge_quality/std': 0.1713601052761078, 'rewards/repeat_penalty/mean': 0.9979227781295776, 'rewards/repeat_penalty/std': 0.005288340616971254, 'rewards/near_duplicate_penalty/mean': 0.9979227781295776, 'rewards/near_duplicate_penalty/std': 0.005288340616971254, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.7766398787498474, 'rewards/total_composite/std': 0.15288196504116058, 'reward': 0.7766398787498474, 'reward_std': 0.15288197994232178, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09629331529140472, 'sampling/sampling_logp_difference/max': 1.5427756309509277, 'sampling/importance_sampling_ratio/min': 0.21378688514232635, 'sampling/importance_sampling_ratio/mean': 1.0054584741592407, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.47175781801342964, 'clip_ratio/low_mean': 0.0070436508394777775, 'clip_ratio/low_min': 0.0070436508394777775, 'clip_ratio/high_mean': 0.063939273590222, 'clip_ratio/high_max': 0.063939273590222, 'clip_ratio/region_mean': 0.07098292442969978, 'reward_total_mean': 0.7766398787498474, 'reward_meter_mean': 0.8920453786849976, 'reward_meter_std': 0.04058147594332695, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.9979227781295776, 'reward_repeat_penalty_std': 0.005288340616971254, 'reward_near_duplicate_penalty_mean': 0.9979227781295776, 'reward_near_duplicate_penalty_std': 0.005288340616971254, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8725000023841858, 'reward_judge_quality_std': 0.1713601052761078, 'reward_total_composite_mean': 0.7766398787498474, 'reward_total_composite_std': 0.15288196504116058, 'epoch': 0.0}

  2%|▏         | 59/3300 [12:50<11:07:25, 12.36s/it]INFO 04-12 10:59:02 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 60/3300 [13:01<10:46:04, 11.96s/it]2026-04-12 10:59:13,105 | INFO | train_grpo_train | metrics_logged mode=train step=60

                                                    
{'loss': 0.111, 'grad_norm': 10.99131965637207, 'learning_rate': 9.821212121212122e-06, 'num_tokens': 134991.0, 'completions/mean_length': 93.625, 'completions/min_length': 68.0, 'completions/max_length': 135.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 93.625, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 135.0, 'rewards/meter/mean': 0.5954556465148926, 'rewards/meter/std': 0.3356790542602539, '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/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.3375000059604645, 'rewards/judge_quality/std': 0.08345229178667068, 'rewards/repeat_penalty/mean': 0.9892946481704712, 'rewards/repeat_penalty/std': 0.0107206329703331, 'rewards/near_duplicate_penalty/mean': 0.9929076433181763, 'rewards/near_duplicate_penalty/std': 0.005824347026646137, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9963715672492981, 'rewards/distinct_2/std': 0.010262799449265003, 'rewards/total_composite/mean': 0.17966029047966003, 'rewards/total_composite/std': 0.10174994170665741, 'reward': 0.17966029047966003, 'reward_std': 0.10174994170665741, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21685421466827393, 'sampling/sampling_logp_difference/max': 1.3862667083740234, 'sampling/importance_sampling_ratio/min': 0.25000691413879395, 'sampling/importance_sampling_ratio/mean': 1.0362759828567505, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6094531416893005, 'clip_ratio/low_mean': 0.09689009841531515, 'clip_ratio/low_min': 0.09689009841531515, 'clip_ratio/high_mean': 0.08354575373232365, 'clip_ratio/high_max': 0.08354575373232365, 'clip_ratio/region_mean': 0.1804358521476388, 'reward_total_mean': 0.17966029047966003, 'reward_meter_mean': 0.5954556465148926, 'reward_meter_std': 0.3356790542602539, '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_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.9892946481704712, 'reward_repeat_penalty_std': 0.0107206329703331, 'reward_near_duplicate_penalty_mean': 0.9929076433181763, 'reward_near_duplicate_penalty_std': 0.005824347026646137, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9963715672492981, 'reward_distinct_2_std': 0.010262799449265003, 'reward_judge_quality_mean': 0.3375000059604645, 'reward_judge_quality_std': 0.08345229178667068, 'reward_total_composite_mean': 0.17966029047966003, 'reward_total_composite_std': 0.10174994170665741, 'epoch': 0.0}

  2%|▏         | 60/3300 [13:01<10:46:04, 11.96s/it]INFO 04-12 10:59:13 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 61/3300 [13:11<10:22:08, 11.52s/it]2026-04-12 10:59:23,602 | INFO | train_grpo_train | metrics_logged mode=train step=61

                                                    
{'loss': -0.0201, 'grad_norm': 11.589716911315918, 'learning_rate': 9.81818181818182e-06, 'num_tokens': 136849.0, 'completions/mean_length': 67.25, 'completions/min_length': 49.0, 'completions/max_length': 91.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.25, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 91.0, 'rewards/meter/mean': 0.26395222544670105, 'rewards/meter/std': 0.23474130034446716, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9916572570800781, 'rewards/lexical_plausibility/std': 0.023596882820129395, 'rewards/judge_quality/mean': 0.44999998807907104, 'rewards/judge_quality/std': 0.13093073666095734, 'rewards/repeat_penalty/mean': 0.9916759133338928, 'rewards/repeat_penalty/std': 0.008064454421401024, 'rewards/near_duplicate_penalty/mean': 0.9916759133338928, 'rewards/near_duplicate_penalty/std': 0.008064454421401024, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.12356916815042496, 'rewards/total_composite/std': 0.10796519368886948, 'reward': 0.12356916815042496, 'reward_std': 0.10796519368886948, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21231819689273834, 'sampling/sampling_logp_difference/max': 1.8033027648925781, 'sampling/importance_sampling_ratio/min': 0.16475383937358856, 'sampling/importance_sampling_ratio/mean': 1.039514183998108, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9188917577266693, 'clip_ratio/low_mean': 0.11141200363636017, 'clip_ratio/low_min': 0.11141200363636017, 'clip_ratio/high_mean': 0.10190850868821144, 'clip_ratio/high_max': 0.10190850868821144, 'clip_ratio/region_mean': 0.2133205123245716, 'reward_total_mean': 0.12356916815042496, 'reward_meter_mean': 0.26395222544670105, 'reward_meter_std': 0.23474130034446716, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9916572570800781, 'reward_lexical_plausibility_std': 0.023596882820129395, 'reward_repeat_penalty_mean': 0.9916759133338928, 'reward_repeat_penalty_std': 0.008064454421401024, 'reward_near_duplicate_penalty_mean': 0.9916759133338928, 'reward_near_duplicate_penalty_std': 0.008064454421401024, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.44999998807907104, 'reward_judge_quality_std': 0.13093073666095734, 'reward_total_composite_mean': 0.12356916815042496, 'reward_total_composite_std': 0.10796519368886948, 'epoch': 0.0}

  2%|▏         | 61/3300 [13:11<10:22:08, 11.52s/it]INFO 04-12 10:59:23 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 62/3300 [13:22<10:15:28, 11.40s/it]2026-04-12 10:59:34,728 | INFO | train_grpo_train | metrics_logged mode=train step=62

                                                    
{'loss': 0.0066, 'grad_norm': 8.685428619384766, 'learning_rate': 9.815151515151516e-06, 'num_tokens': 139273.0, 'completions/mean_length': 138.0, 'completions/min_length': 85.0, 'completions/max_length': 189.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 138.0, 'completions/min_terminated_length': 85.0, 'completions/max_terminated_length': 189.0, 'rewards/meter/mean': 0.4546557664871216, 'rewards/meter/std': 0.29831260442733765, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.13363061845302582, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9986978769302368, 'rewards/lexical_plausibility/std': 0.0036828548181802034, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.08345229178667068, 'rewards/repeat_penalty/mean': 0.9953555464744568, 'rewards/repeat_penalty/std': 0.004151875153183937, 'rewards/near_duplicate_penalty/mean': 0.9953555464744568, 'rewards/near_duplicate_penalty/std': 0.004151875153183937, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.11611887067556381, 'rewards/total_composite/std': 0.083598792552948, 'reward': 0.11611887067556381, 'reward_std': 0.083598792552948, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21783751249313354, 'sampling/sampling_logp_difference/max': 1.8963346481323242, 'sampling/importance_sampling_ratio/min': 0.15011785924434662, 'sampling/importance_sampling_ratio/mean': 1.040421485900879, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.382453352212906, 'clip_ratio/low_mean': 0.1357566211372614, 'clip_ratio/low_min': 0.1357566211372614, 'clip_ratio/high_mean': 0.05143502727150917, 'clip_ratio/high_max': 0.05143502727150917, 'clip_ratio/region_mean': 0.18719164840877056, 'reward_total_mean': 0.11611887067556381, 'reward_meter_mean': 0.4546557664871216, 'reward_meter_std': 0.29831260442733765, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.13363061845302582, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9986978769302368, 'reward_lexical_plausibility_std': 0.0036828548181802034, 'reward_repeat_penalty_mean': 0.9953555464744568, 'reward_repeat_penalty_std': 0.004151875153183937, 'reward_near_duplicate_penalty_mean': 0.9953555464744568, 'reward_near_duplicate_penalty_std': 0.004151875153183937, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.08345229178667068, 'reward_total_composite_mean': 0.11611887067556381, 'reward_total_composite_std': 0.083598792552948, 'epoch': 0.0}

  2%|▏         | 62/3300 [13:22<10:15:28, 11.40s/it]INFO 04-12 10:59:34 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 63/3300 [13:31<9:37:10, 10.70s/it] 2026-04-12 10:59:43,778 | INFO | train_grpo_train | metrics_logged mode=train step=63

                                                   
{'loss': 0.0949, 'grad_norm': 18.836515426635742, 'learning_rate': 9.812121212121212e-06, 'num_tokens': 140891.0, 'completions/mean_length': 39.25, 'completions/min_length': 31.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.25, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.5631728172302246, 'rewards/meter/std': 0.45356348156929016, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9866071343421936, 'rewards/lexical_plausibility/std': 0.03788072615861893, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.16903084516525269, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1947702318429947, 'rewards/total_composite/std': 0.20791685581207275, 'reward': 0.1947702318429947, 'reward_std': 0.20791685581207275, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2171783745288849, 'sampling/sampling_logp_difference/max': 1.5032119750976562, 'sampling/importance_sampling_ratio/min': 0.22241462767124176, 'sampling/importance_sampling_ratio/mean': 1.0317139625549316, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.506786197423935, 'clip_ratio/low_mean': 0.1136647928506136, 'clip_ratio/low_min': 0.1136647928506136, 'clip_ratio/high_mean': 0.09549001883715391, 'clip_ratio/high_max': 0.09549001883715391, 'clip_ratio/region_mean': 0.2091548116877675, 'reward_total_mean': 0.1947702318429947, 'reward_meter_mean': 0.5631728172302246, 'reward_meter_std': 0.45356348156929016, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9866071343421936, 'reward_lexical_plausibility_std': 0.03788072615861893, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.16903084516525269, 'reward_total_composite_mean': 0.1947702318429947, 'reward_total_composite_std': 0.20791685581207275, 'epoch': 0.0}

  2%|▏         | 63/3300 [13:31<9:37:10, 10.70s/it]INFO 04-12 10:59:44 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 64/3300 [13:42<9:30:51, 10.58s/it]2026-04-12 10:59:54,097 | INFO | train_grpo_train | metrics_logged mode=train step=64

                                                   
{'loss': 0.1169, 'grad_norm': 7.857104301452637, 'learning_rate': 9.809090909090911e-06, 'num_tokens': 143499.0, 'completions/mean_length': 141.0, 'completions/min_length': 116.0, 'completions/max_length': 176.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 141.0, 'completions/min_terminated_length': 116.0, 'completions/max_terminated_length': 176.0, 'rewards/meter/mean': 0.8880764245986938, 'rewards/meter/std': 0.2671806514263153, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.13363061845302582, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9154520630836487, 'rewards/lexical_plausibility/std': 0.11608117818832397, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.10690449178218842, 'rewards/repeat_penalty/mean': 0.9962418079376221, 'rewards/repeat_penalty/std': 0.0031225020065903664, 'rewards/near_duplicate_penalty/mean': 0.9962418079376221, 'rewards/near_duplicate_penalty/std': 0.0031225020065903664, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1832917481660843, 'rewards/total_composite/std': 0.09190372377634048, 'reward': 0.1832917481660843, 'reward_std': 0.09190372377634048, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20457211136817932, 'sampling/sampling_logp_difference/max': 2.1627116203308105, 'sampling/importance_sampling_ratio/min': 0.11501282453536987, 'sampling/importance_sampling_ratio/mean': 1.0372601747512817, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6322957575321198, 'clip_ratio/low_mean': 0.1129800220951438, 'clip_ratio/low_min': 0.1129800220951438, 'clip_ratio/high_mean': 0.08206339180469513, 'clip_ratio/high_max': 0.08206339180469513, 'clip_ratio/region_mean': 0.19504341389983892, 'reward_total_mean': 0.1832917481660843, 'reward_meter_mean': 0.8880764245986938, 'reward_meter_std': 0.2671806514263153, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.13363061845302582, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9154520630836487, 'reward_lexical_plausibility_std': 0.11608117818832397, 'reward_repeat_penalty_mean': 0.9962418079376221, 'reward_repeat_penalty_std': 0.0031225020065903664, 'reward_near_duplicate_penalty_mean': 0.9962418079376221, 'reward_near_duplicate_penalty_std': 0.0031225020065903664, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.10690449178218842, 'reward_total_composite_mean': 0.1832917481660843, 'reward_total_composite_std': 0.09190372377634048, 'epoch': 0.0}

  2%|▏         | 64/3300 [13:42<9:30:51, 10.58s/it]INFO 04-12 10:59:54 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 65/3300 [13:53<9:42:55, 10.81s/it]2026-04-12 11:00:05,438 | INFO | train_grpo_train | metrics_logged mode=train step=65

                                                   
{'loss': 0.138, 'grad_norm': 11.104138374328613, 'learning_rate': 9.806060606060607e-06, 'num_tokens': 145773.0, 'completions/mean_length': 91.25, 'completions/min_length': 58.0, 'completions/max_length': 118.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 91.25, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 118.0, 'rewards/meter/mean': 0.5430281758308411, 'rewards/meter/std': 0.4575187861919403, '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/lexical_plausibility/mean': 0.9807285666465759, 'rewards/lexical_plausibility/std': 0.054507844150066376, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/repeat_penalty/mean': 0.9975701570510864, 'rewards/repeat_penalty/std': 0.0033119902946054935, 'rewards/near_duplicate_penalty/mean': 0.9975701570510864, 'rewards/near_duplicate_penalty/std': 0.0033119902946054935, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1771267056465149, 'rewards/total_composite/std': 0.17245365679264069, 'reward': 0.1771267056465149, 'reward_std': 0.1724536418914795, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19685818254947662, 'sampling/sampling_logp_difference/max': 1.9196500778198242, 'sampling/importance_sampling_ratio/min': 0.1466582715511322, 'sampling/importance_sampling_ratio/mean': 1.0336054563522339, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.221826910972595, 'clip_ratio/low_mean': 0.09496940113604069, 'clip_ratio/low_min': 0.09496940113604069, 'clip_ratio/high_mean': 0.06819553300738335, 'clip_ratio/high_max': 0.06819553300738335, 'clip_ratio/region_mean': 0.16316493414342403, 'reward_total_mean': 0.1771267056465149, 'reward_meter_mean': 0.5430281758308411, 'reward_meter_std': 0.4575187861919403, '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_lexical_plausibility_mean': 0.9807285666465759, 'reward_lexical_plausibility_std': 0.054507844150066376, 'reward_repeat_penalty_mean': 0.9975701570510864, 'reward_repeat_penalty_std': 0.0033119902946054935, 'reward_near_duplicate_penalty_mean': 0.9975701570510864, 'reward_near_duplicate_penalty_std': 0.0033119902946054935, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.1771267056465149, 'reward_total_composite_std': 0.17245365679264069, 'epoch': 0.0}

  2%|▏         | 65/3300 [13:53<9:42:55, 10.81s/it]INFO 04-12 11:00:05 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 66/3300 [14:03<9:24:38, 10.48s/it]2026-04-12 11:00:15,131 | INFO | train_grpo_train | metrics_logged mode=train step=66

                                                   
{'loss': 0.0022, 'grad_norm': 24.601667404174805, 'learning_rate': 9.803030303030304e-06, 'num_tokens': 147548.0, 'completions/mean_length': 41.875, 'completions/min_length': 29.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.875, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.978650689125061, 'rewards/meter/std': 0.05017535015940666, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8212499618530273, 'rewards/judge_quality/std': 0.22937415540218353, 'rewards/repeat_penalty/mean': 0.9656137824058533, 'rewards/repeat_penalty/std': 0.0294276624917984, 'rewards/near_duplicate_penalty/mean': 0.9656137824058533, 'rewards/near_duplicate_penalty/std': 0.0294276624917984, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.8102796077728271, 'rewards/total_composite/std': 0.24384412169456482, 'reward': 0.8102796077728271, 'reward_std': 0.24384410679340363, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1336280256509781, 'sampling/sampling_logp_difference/max': 2.570559501647949, 'sampling/importance_sampling_ratio/min': 0.07649273425340652, 'sampling/importance_sampling_ratio/mean': 0.9901538491249084, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.46303134225308895, 'clip_ratio/low_mean': 0.04460171051323414, 'clip_ratio/low_min': 0.04460171051323414, 'clip_ratio/high_mean': 0.05560530349612236, 'clip_ratio/high_max': 0.05560530349612236, 'clip_ratio/region_mean': 0.1002070140093565, 'reward_total_mean': 0.8102796077728271, 'reward_meter_mean': 0.978650689125061, 'reward_meter_std': 0.05017535015940666, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9656137824058533, 'reward_repeat_penalty_std': 0.0294276624917984, 'reward_near_duplicate_penalty_mean': 0.9656137824058533, 'reward_near_duplicate_penalty_std': 0.0294276624917984, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8212499618530273, 'reward_judge_quality_std': 0.22937415540218353, 'reward_total_composite_mean': 0.8102796077728271, 'reward_total_composite_std': 0.24384412169456482, 'epoch': 0.0}

  2%|▏         | 66/3300 [14:03<9:24:38, 10.48s/it]INFO 04-12 11:00:15 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 67/3300 [14:12<9:04:36, 10.11s/it]2026-04-12 11:00:24,378 | INFO | train_grpo_train | metrics_logged mode=train step=67

                                                   
{'loss': 0.023, 'grad_norm': 20.278247833251953, 'learning_rate': 9.800000000000001e-06, 'num_tokens': 149152.0, 'completions/mean_length': 42.5, 'completions/min_length': 30.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.5, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.5701628923416138, 'rewards/meter/std': 0.36496612429618835, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9910990595817566, 'rewards/lexical_plausibility/std': 0.02517561987042427, 'rewards/judge_quality/mean': 0.5549999475479126, 'rewards/judge_quality/std': 0.22790977358818054, 'rewards/repeat_penalty/mean': 0.9709295034408569, 'rewards/repeat_penalty/std': 0.03340797498822212, 'rewards/near_duplicate_penalty/mean': 0.9709295034408569, 'rewards/near_duplicate_penalty/std': 0.03340797498822212, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.32349228858947754, 'rewards/total_composite/std': 0.25729161500930786, 'reward': 0.32349228858947754, 'reward_std': 0.25729164481163025, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22895435988903046, 'sampling/sampling_logp_difference/max': 1.4766225814819336, 'sampling/importance_sampling_ratio/min': 0.2284078299999237, 'sampling/importance_sampling_ratio/mean': 1.0299285650253296, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8747659176588058, 'clip_ratio/low_mean': 0.1095969807356596, 'clip_ratio/low_min': 0.1095969807356596, 'clip_ratio/high_mean': 0.0692050140351057, 'clip_ratio/high_max': 0.0692050140351057, 'clip_ratio/region_mean': 0.1788019947707653, 'reward_total_mean': 0.32349228858947754, 'reward_meter_mean': 0.5701628923416138, 'reward_meter_std': 0.36496612429618835, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9910990595817566, 'reward_lexical_plausibility_std': 0.02517561987042427, 'reward_repeat_penalty_mean': 0.9709295034408569, 'reward_repeat_penalty_std': 0.03340797498822212, 'reward_near_duplicate_penalty_mean': 0.9709295034408569, 'reward_near_duplicate_penalty_std': 0.03340797498822212, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5549999475479126, 'reward_judge_quality_std': 0.22790977358818054, 'reward_total_composite_mean': 0.32349228858947754, 'reward_total_composite_std': 0.25729161500930786, 'epoch': 0.0}

  2%|▏         | 67/3300 [14:12<9:04:36, 10.11s/it]INFO 04-12 11:00:24 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 68/3300 [14:22<8:57:41,  9.98s/it]2026-04-12 11:00:34,068 | INFO | train_grpo_train | metrics_logged mode=train step=68

                                                   
{'loss': -0.0168, 'grad_norm': 16.09491729736328, 'learning_rate': 9.796969696969698e-06, 'num_tokens': 150820.0, 'completions/mean_length': 53.5, 'completions/min_length': 35.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.5, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.37656503915786743, 'rewards/meter/std': 0.3850724697113037, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9549214839935303, 'rewards/lexical_plausibility/std': 0.10456041991710663, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.1060660183429718, 'rewards/repeat_penalty/mean': 0.9830353260040283, 'rewards/repeat_penalty/std': 0.024728700518608093, 'rewards/near_duplicate_penalty/mean': 0.9830353260040283, 'rewards/near_duplicate_penalty/std': 0.024728700518608093, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1273587942123413, 'rewards/total_composite/std': 0.13430751860141754, 'reward': 0.1273587942123413, 'reward_std': 0.13430751860141754, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20365306735038757, 'sampling/sampling_logp_difference/max': 1.881567358970642, 'sampling/importance_sampling_ratio/min': 0.15235112607479095, 'sampling/importance_sampling_ratio/mean': 1.0225739479064941, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.723569542169571, 'clip_ratio/low_mean': 0.1031299326568842, 'clip_ratio/low_min': 0.1031299326568842, 'clip_ratio/high_mean': 0.11396108567714691, 'clip_ratio/high_max': 0.11396108567714691, 'clip_ratio/region_mean': 0.2170910183340311, 'reward_total_mean': 0.1273587942123413, 'reward_meter_mean': 0.37656503915786743, 'reward_meter_std': 0.3850724697113037, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9549214839935303, 'reward_lexical_plausibility_std': 0.10456041991710663, 'reward_repeat_penalty_mean': 0.9830353260040283, 'reward_repeat_penalty_std': 0.024728700518608093, 'reward_near_duplicate_penalty_mean': 0.9830353260040283, 'reward_near_duplicate_penalty_std': 0.024728700518608093, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.1060660183429718, 'reward_total_composite_mean': 0.1273587942123413, 'reward_total_composite_std': 0.13430751860141754, 'epoch': 0.0}

  2%|▏         | 68/3300 [14:22<8:57:41,  9.98s/it]INFO 04-12 11:00:34 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 69/3300 [14:29<8:19:28,  9.28s/it]2026-04-12 11:00:41,699 | INFO | train_grpo_train | metrics_logged mode=train step=69

                                                   
{'loss': -0.018, 'grad_norm': 40.95964050292969, 'learning_rate': 9.793939393939394e-06, 'num_tokens': 152206.0, 'completions/mean_length': 25.25, 'completions/min_length': 22.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.25, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.8356223702430725, 'rewards/meter/std': 0.24600468575954437, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.9424999952316284, 'rewards/judge_quality/std': 0.013887288980185986, 'rewards/repeat_penalty/mean': 0.9844276309013367, 'rewards/repeat_penalty/std': 0.00961147528141737, 'rewards/near_duplicate_penalty/mean': 0.9844276309013367, 'rewards/near_duplicate_penalty/std': 0.00961147528141737, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.7894231081008911, 'rewards/total_composite/std': 0.23446962237358093, 'reward': 0.7894231081008911, 'reward_std': 0.23446960747241974, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07563488930463791, 'sampling/sampling_logp_difference/max': 1.2072997093200684, 'sampling/importance_sampling_ratio/min': 0.29900357127189636, 'sampling/importance_sampling_ratio/mean': 0.9926040768623352, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.23366530891507864, 'clip_ratio/low_mean': 0.005681818351149559, 'clip_ratio/low_min': 0.005681818351149559, 'clip_ratio/high_mean': 0.06956934370100498, 'clip_ratio/high_max': 0.06956934370100498, 'clip_ratio/region_mean': 0.07525116205215454, 'reward_total_mean': 0.7894231081008911, 'reward_meter_mean': 0.8356223702430725, 'reward_meter_std': 0.24600468575954437, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9844276309013367, 'reward_repeat_penalty_std': 0.00961147528141737, 'reward_near_duplicate_penalty_mean': 0.9844276309013367, 'reward_near_duplicate_penalty_std': 0.00961147528141737, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.9424999952316284, 'reward_judge_quality_std': 0.013887288980185986, 'reward_total_composite_mean': 0.7894231081008911, 'reward_total_composite_std': 0.23446962237358093, 'epoch': 0.0}

  2%|▏         | 69/3300 [14:29<8:19:28,  9.28s/it]INFO 04-12 11:00:42 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 70/3300 [14:44<9:47:24, 10.91s/it]2026-04-12 11:00:56,424 | INFO | train_grpo_train | metrics_logged mode=train step=70

                                                   
{'loss': -0.1946, 'grad_norm': 3.6259946823120117, 'learning_rate': 9.790909090909093e-06, 'num_tokens': 156279.0, 'completions/mean_length': 417.125, 'completions/min_length': 310.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 385.5, 'completions/min_terminated_length': 310.0, 'completions/max_terminated_length': 502.0, 'rewards/meter/mean': 0.8855130672454834, 'rewards/meter/std': 0.10192851722240448, 'rewards/count_adherence/mean': 0.9038461446762085, 'rewards/count_adherence/std': 0.03560846298933029, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/repeat_penalty/mean': 0.9922628998756409, 'rewards/repeat_penalty/std': 0.005649227183312178, 'rewards/near_duplicate_penalty/mean': 0.9961462020874023, 'rewards/near_duplicate_penalty/std': 0.0011077204253524542, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9961007833480835, 'rewards/distinct_2/std': 0.005382579285651445, 'rewards/total_composite/mean': 0.2245791256427765, 'rewards/total_composite/std': 0.13737398386001587, 'reward': 0.2245791256427765, 'reward_std': 0.13737396895885468, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2077775001525879, 'sampling/sampling_logp_difference/max': 1.387526512145996, 'sampling/importance_sampling_ratio/min': 0.2496921569108963, 'sampling/importance_sampling_ratio/mean': 1.040596604347229, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.358903408050537, 'clip_ratio/low_mean': 0.020779220387339592, 'clip_ratio/low_min': 0.020779220387339592, 'clip_ratio/high_mean': 0.12808114290237427, 'clip_ratio/high_max': 0.12808114290237427, 'clip_ratio/region_mean': 0.14886036328971386, 'reward_total_mean': 0.2245791256427765, 'reward_meter_mean': 0.8855130672454834, 'reward_meter_std': 0.10192851722240448, 'reward_count_adherence_mean': 0.9038461446762085, 'reward_count_adherence_std': 0.03560846298933029, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9922628998756409, 'reward_repeat_penalty_std': 0.005649227183312178, 'reward_near_duplicate_penalty_mean': 0.9961462020874023, 'reward_near_duplicate_penalty_std': 0.0011077204253524542, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9961007833480835, 'reward_distinct_2_std': 0.005382579285651445, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.2245791256427765, 'reward_total_composite_std': 0.13737398386001587, 'epoch': 0.0}

  2%|▏         | 70/3300 [14:44<9:47:24, 10.91s/it]INFO 04-12 11:00:56 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 71/3300 [14:53<9:14:23, 10.30s/it]2026-04-12 11:01:05,301 | INFO | train_grpo_train | metrics_logged mode=train step=71

                                                   
{'loss': 0.0743, 'grad_norm': 20.082340240478516, 'learning_rate': 9.787878787878788e-06, 'num_tokens': 157806.0, 'completions/mean_length': 31.875, 'completions/min_length': 25.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.875, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.7298133969306946, 'rewards/meter/std': 0.34841012954711914, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6762499809265137, 'rewards/judge_quality/std': 0.2707628309726715, 'rewards/repeat_penalty/mean': 0.8956765532493591, 'rewards/repeat_penalty/std': 0.10391052812337875, 'rewards/near_duplicate_penalty/mean': 0.987528920173645, 'rewards/near_duplicate_penalty/std': 0.008463931269943714, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9065170884132385, 'rewards/distinct_2/std': 0.10018188506364822, 'rewards/total_composite/mean': 0.5111167430877686, 'rewards/total_composite/std': 0.33344566822052, 'reward': 0.5111167430877686, 'reward_std': 0.33344566822052, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16924262046813965, 'sampling/sampling_logp_difference/max': 1.339747428894043, 'sampling/importance_sampling_ratio/min': 0.2619118094444275, 'sampling/importance_sampling_ratio/mean': 0.9970604181289673, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1145559754222631, 'clip_ratio/low_mean': 0.1381722390651703, 'clip_ratio/low_min': 0.1381722390651703, 'clip_ratio/high_mean': 0.011363636702299118, 'clip_ratio/high_max': 0.011363636702299118, 'clip_ratio/region_mean': 0.1495358757674694, 'reward_total_mean': 0.5111167430877686, 'reward_meter_mean': 0.7298133969306946, 'reward_meter_std': 0.34841012954711914, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8956765532493591, 'reward_repeat_penalty_std': 0.10391052812337875, 'reward_near_duplicate_penalty_mean': 0.987528920173645, 'reward_near_duplicate_penalty_std': 0.008463931269943714, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9065170884132385, 'reward_distinct_2_std': 0.10018188506364822, 'reward_judge_quality_mean': 0.6762499809265137, 'reward_judge_quality_std': 0.2707628309726715, 'reward_total_composite_mean': 0.5111167430877686, 'reward_total_composite_std': 0.33344566822052, 'epoch': 0.0}

  2%|▏         | 71/3300 [14:53<9:14:23, 10.30s/it]INFO 04-12 11:01:05 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 72/3300 [15:02<9:02:40, 10.09s/it]2026-04-12 11:01:14,888 | INFO | train_grpo_train | metrics_logged mode=train step=72

                                                   
{'loss': -0.0497, 'grad_norm': 14.805048942565918, 'learning_rate': 9.784848484848486e-06, 'num_tokens': 159635.0, 'completions/mean_length': 52.625, 'completions/min_length': 36.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.625, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.45533621311187744, 'rewards/meter/std': 0.4449755549430847, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.9885790348052979, 'rewards/repeat_penalty/std': 0.031295377761125565, 'rewards/near_duplicate_penalty/mean': 0.9986138343811035, 'rewards/near_duplicate_penalty/std': 0.0030318740755319595, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9898785352706909, 'rewards/distinct_2/std': 0.028627805411815643, 'rewards/total_composite/mean': 0.13843142986297607, 'rewards/total_composite/std': 0.13367311656475067, 'reward': 0.13843142986297607, 'reward_std': 0.13367311656475067, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2201484739780426, 'sampling/sampling_logp_difference/max': 1.6301121711730957, 'sampling/importance_sampling_ratio/min': 0.1959075927734375, 'sampling/importance_sampling_ratio/mean': 1.0505764484405518, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2289724200963974, 'clip_ratio/low_mean': 0.09259190503507853, 'clip_ratio/low_min': 0.09259190503507853, 'clip_ratio/high_mean': 0.10724222846329212, 'clip_ratio/high_max': 0.10724222846329212, 'clip_ratio/region_mean': 0.19983413349837065, 'reward_total_mean': 0.13843142986297607, 'reward_meter_mean': 0.45533621311187744, 'reward_meter_std': 0.4449755549430847, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.9885790348052979, 'reward_repeat_penalty_std': 0.031295377761125565, 'reward_near_duplicate_penalty_mean': 0.9986138343811035, 'reward_near_duplicate_penalty_std': 0.0030318740755319595, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9898785352706909, 'reward_distinct_2_std': 0.028627805411815643, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.13843142986297607, 'reward_total_composite_std': 0.13367311656475067, 'epoch': 0.0}

  2%|▏         | 72/3300 [15:02<9:02:40, 10.09s/it]INFO 04-12 11:01:15 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 73/3300 [15:11<8:45:19,  9.77s/it]2026-04-12 11:01:23,910 | INFO | train_grpo_train | metrics_logged mode=train step=73

                                                   
{'loss': 0.0899, 'grad_norm': 17.469419479370117, 'learning_rate': 9.781818181818183e-06, 'num_tokens': 161138.0, 'completions/mean_length': 37.875, 'completions/min_length': 29.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.875, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.7943594455718994, 'rewards/meter/std': 0.30483105778694153, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.10350983589887619, 'rewards/repeat_penalty/mean': 0.9412729740142822, 'rewards/repeat_penalty/std': 0.1250309944152832, 'rewards/near_duplicate_penalty/mean': 0.9755563735961914, 'rewards/near_duplicate_penalty/std': 0.02905178628861904, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9912587404251099, 'rewards/distinct_2/std': 0.024724015966057777, 'rewards/total_composite/mean': 0.28405147790908813, 'rewards/total_composite/std': 0.1237814724445343, 'reward': 0.28405147790908813, 'reward_std': 0.1237814724445343, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2270776480436325, 'sampling/sampling_logp_difference/max': 1.622610092163086, 'sampling/importance_sampling_ratio/min': 0.197382852435112, 'sampling/importance_sampling_ratio/mean': 1.0431281328201294, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2778522223234177, 'clip_ratio/low_mean': 0.06417410634458065, 'clip_ratio/low_min': 0.06417410634458065, 'clip_ratio/high_mean': 0.14028591848909855, 'clip_ratio/high_max': 0.14028591848909855, 'clip_ratio/region_mean': 0.2044600248336792, 'reward_total_mean': 0.28405147790908813, 'reward_meter_mean': 0.7943594455718994, 'reward_meter_std': 0.30483105778694153, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9412729740142822, 'reward_repeat_penalty_std': 0.1250309944152832, 'reward_near_duplicate_penalty_mean': 0.9755563735961914, 'reward_near_duplicate_penalty_std': 0.02905178628861904, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9912587404251099, 'reward_distinct_2_std': 0.024724015966057777, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.10350983589887619, 'reward_total_composite_mean': 0.28405147790908813, 'reward_total_composite_std': 0.1237814724445343, 'epoch': 0.0}

  2%|▏         | 73/3300 [15:11<8:45:19,  9.77s/it]INFO 04-12 11:01:24 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 74/3300 [15:20<8:26:27,  9.42s/it]2026-04-12 11:01:32,518 | INFO | train_grpo_train | metrics_logged mode=train step=74

                                                   
{'loss': 0.043, 'grad_norm': 15.154500007629395, 'learning_rate': 9.77878787878788e-06, 'num_tokens': 162839.0, 'completions/mean_length': 44.625, 'completions/min_length': 37.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.625, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.6420912146568298, 'rewards/meter/std': 0.3212152421474457, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.5337499976158142, 'rewards/judge_quality/std': 0.23445606231689453, 'rewards/repeat_penalty/mean': 0.9913415908813477, 'rewards/repeat_penalty/std': 0.014010717160999775, 'rewards/near_duplicate_penalty/mean': 0.9913415908813477, 'rewards/near_duplicate_penalty/std': 0.014010717160999775, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3114110231399536, 'rewards/total_composite/std': 0.16819478571414948, 'reward': 0.3114110231399536, 'reward_std': 0.16819478571414948, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22432444989681244, 'sampling/sampling_logp_difference/max': 1.4277992248535156, 'sampling/importance_sampling_ratio/min': 0.239836186170578, 'sampling/importance_sampling_ratio/mean': 1.0319697856903076, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8442463725805283, 'clip_ratio/low_mean': 0.08643845655024052, 'clip_ratio/low_min': 0.08643845655024052, 'clip_ratio/high_mean': 0.12991708144545555, 'clip_ratio/high_max': 0.12991708144545555, 'clip_ratio/region_mean': 0.21635553799569607, 'reward_total_mean': 0.3114110231399536, 'reward_meter_mean': 0.6420912146568298, 'reward_meter_std': 0.3212152421474457, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.9913415908813477, 'reward_repeat_penalty_std': 0.014010717160999775, 'reward_near_duplicate_penalty_mean': 0.9913415908813477, 'reward_near_duplicate_penalty_std': 0.014010717160999775, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5337499976158142, 'reward_judge_quality_std': 0.23445606231689453, 'reward_total_composite_mean': 0.3114110231399536, 'reward_total_composite_std': 0.16819478571414948, 'epoch': 0.0}

  2%|▏         | 74/3300 [15:20<8:26:27,  9.42s/it]INFO 04-12 11:01:32 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 75/3300 [15:35<10:04:17, 11.24s/it]2026-04-12 11:01:48,014 | INFO | train_grpo_train | metrics_logged mode=train step=75

                                                    
{'loss': 0.2983, 'grad_norm': 3.368990898132324, 'learning_rate': 9.775757575757576e-06, 'num_tokens': 166104.0, 'completions/mean_length': 454.125, 'completions/min_length': 220.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 396.25, 'completions/min_terminated_length': 220.0, 'completions/max_terminated_length': 470.0, 'rewards/meter/mean': 0.6994812488555908, 'rewards/meter/std': 0.17008261382579803, 'rewards/count_adherence/mean': 0.8374999761581421, 'rewards/count_adherence/std': 0.16201850771903992, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9991605877876282, 'rewards/lexical_plausibility/std': 0.0015816990053281188, 'rewards/judge_quality/mean': 0.20000000298023224, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.9920966029167175, 'rewards/repeat_penalty/std': 0.006813052576035261, 'rewards/near_duplicate_penalty/mean': 0.9953973889350891, 'rewards/near_duplicate_penalty/std': 0.0027697717305272818, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9966833591461182, 'rewards/distinct_2/std': 0.006151220761239529, 'rewards/total_composite/mean': 0.09628508985042572, 'rewards/total_composite/std': 0.05001360923051834, 'reward': 0.09628508985042572, 'reward_std': 0.05001360923051834, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22374960780143738, 'sampling/sampling_logp_difference/max': 1.7989654541015625, 'sampling/importance_sampling_ratio/min': 0.16546998918056488, 'sampling/importance_sampling_ratio/mean': 1.0535924434661865, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8021921217441559, 'clip_ratio/low_mean': 0.06391321308910847, 'clip_ratio/low_min': 0.06391321308910847, 'clip_ratio/high_mean': 0.02845744602382183, 'clip_ratio/high_max': 0.02845744602382183, 'clip_ratio/region_mean': 0.0923706591129303, 'reward_total_mean': 0.09628508985042572, 'reward_meter_mean': 0.6994812488555908, 'reward_meter_std': 0.17008261382579803, 'reward_count_adherence_mean': 0.8374999761581421, 'reward_count_adherence_std': 0.16201850771903992, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9991605877876282, 'reward_lexical_plausibility_std': 0.0015816990053281188, 'reward_repeat_penalty_mean': 0.9920966029167175, 'reward_repeat_penalty_std': 0.006813052576035261, 'reward_near_duplicate_penalty_mean': 0.9953973889350891, 'reward_near_duplicate_penalty_std': 0.0027697717305272818, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9966833591461182, 'reward_distinct_2_std': 0.006151220761239529, 'reward_judge_quality_mean': 0.20000000298023224, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.09628508985042572, 'reward_total_composite_std': 0.05001360923051834, 'epoch': 0.0}

  2%|▏         | 75/3300 [15:35<10:04:17, 11.24s/it]INFO 04-12 11:01:48 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 76/3300 [15:46<9:48:41, 10.96s/it] 2026-04-12 11:01:58,301 | INFO | train_grpo_train | metrics_logged mode=train step=76

                                                   
{'loss': 0.155, 'grad_norm': 10.245954513549805, 'learning_rate': 9.772727272727273e-06, 'num_tokens': 168471.0, 'completions/mean_length': 106.875, 'completions/min_length': 68.0, 'completions/max_length': 164.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 106.875, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 164.0, 'rewards/meter/mean': 0.44018733501434326, 'rewards/meter/std': 0.2502119839191437, '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/lexical_plausibility/mean': 0.9872140884399414, 'rewards/lexical_plausibility/std': 0.023682869970798492, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.9958027601242065, 'rewards/repeat_penalty/std': 0.003779177786782384, 'rewards/near_duplicate_penalty/mean': 0.9958027601242065, 'rewards/near_duplicate_penalty/std': 0.003779177786782384, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1325821578502655, 'rewards/total_composite/std': 0.09790375083684921, 'reward': 0.1325821578502655, 'reward_std': 0.09790375083684921, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2237071543931961, 'sampling/sampling_logp_difference/max': 2.673611640930176, 'sampling/importance_sampling_ratio/min': 0.06900256127119064, 'sampling/importance_sampling_ratio/mean': 1.0415984392166138, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.82192063331604, 'clip_ratio/low_mean': 0.10153472051024437, 'clip_ratio/low_min': 0.10153472051024437, 'clip_ratio/high_mean': 0.06862385757267475, 'clip_ratio/high_max': 0.06862385757267475, 'clip_ratio/region_mean': 0.17015857808291912, 'reward_total_mean': 0.1325821578502655, 'reward_meter_mean': 0.44018733501434326, 'reward_meter_std': 0.2502119839191437, '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_lexical_plausibility_mean': 0.9872140884399414, 'reward_lexical_plausibility_std': 0.023682869970798492, 'reward_repeat_penalty_mean': 0.9958027601242065, 'reward_repeat_penalty_std': 0.003779177786782384, 'reward_near_duplicate_penalty_mean': 0.9958027601242065, 'reward_near_duplicate_penalty_std': 0.003779177786782384, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.1325821578502655, 'reward_total_composite_std': 0.09790375083684921, 'epoch': 0.0}

  2%|▏         | 76/3300 [15:46<9:48:41, 10.96s/it]INFO 04-12 11:01:58 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 77/3300 [15:57<9:48:46, 10.96s/it]2026-04-12 11:02:09,274 | INFO | train_grpo_train | metrics_logged mode=train step=77

                                                   
{'loss': 0.0214, 'grad_norm': 12.224149703979492, 'learning_rate': 9.76969696969697e-06, 'num_tokens': 170378.0, 'completions/mean_length': 74.375, 'completions/min_length': 53.0, 'completions/max_length': 102.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.375, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.5974334478378296, 'rewards/meter/std': 0.4106157124042511, '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/lexical_plausibility/mean': 0.9610192775726318, 'rewards/lexical_plausibility/std': 0.09549286216497421, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.1060660183429718, 'rewards/repeat_penalty/mean': 0.9934601187705994, 'rewards/repeat_penalty/std': 0.005460478365421295, 'rewards/near_duplicate_penalty/mean': 0.9934601187705994, 'rewards/near_duplicate_penalty/std': 0.005460478365421295, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.24743841588497162, 'rewards/total_composite/std': 0.20405541360378265, 'reward': 0.24743841588497162, 'reward_std': 0.20405542850494385, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2201482504606247, 'sampling/sampling_logp_difference/max': 1.216226577758789, 'sampling/importance_sampling_ratio/min': 0.2963462769985199, 'sampling/importance_sampling_ratio/mean': 1.0458892583847046, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.8140107095241547, 'clip_ratio/low_mean': 0.10759576968848705, 'clip_ratio/low_min': 0.10759576968848705, 'clip_ratio/high_mean': 0.10649478435516357, 'clip_ratio/high_max': 0.10649478435516357, 'clip_ratio/region_mean': 0.21409055404365063, 'reward_total_mean': 0.24743841588497162, 'reward_meter_mean': 0.5974334478378296, 'reward_meter_std': 0.4106157124042511, '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_lexical_plausibility_mean': 0.9610192775726318, 'reward_lexical_plausibility_std': 0.09549286216497421, 'reward_repeat_penalty_mean': 0.9934601187705994, 'reward_repeat_penalty_std': 0.005460478365421295, 'reward_near_duplicate_penalty_mean': 0.9934601187705994, 'reward_near_duplicate_penalty_std': 0.005460478365421295, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.1060660183429718, 'reward_total_composite_mean': 0.24743841588497162, 'reward_total_composite_std': 0.20405541360378265, 'epoch': 0.0}

  2%|▏         | 77/3300 [15:57<9:48:46, 10.96s/it]INFO 04-12 11:02:09 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 78/3300 [16:08<9:51:21, 11.01s/it]2026-04-12 11:02:20,406 | INFO | train_grpo_train | metrics_logged mode=train step=78

                                                   
{'loss': 0.0933, 'grad_norm': 7.355525493621826, 'learning_rate': 9.766666666666667e-06, 'num_tokens': 173461.0, 'completions/mean_length': 179.375, 'completions/min_length': 152.0, 'completions/max_length': 205.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 179.375, 'completions/min_terminated_length': 152.0, 'completions/max_terminated_length': 205.0, 'rewards/meter/mean': 0.7604238390922546, 'rewards/meter/std': 0.27199587225914, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3375000059604645, 'rewards/judge_quality/std': 0.08345229178667068, 'rewards/repeat_penalty/mean': 0.9908215999603271, 'rewards/repeat_penalty/std': 0.009297901764512062, 'rewards/near_duplicate_penalty/mean': 0.9939049482345581, 'rewards/near_duplicate_penalty/std': 0.0034192020539194345, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9968982934951782, 'rewards/distinct_2/std': 0.008773035369813442, 'rewards/total_composite/mean': 0.20494306087493896, 'rewards/total_composite/std': 0.0963810607790947, 'reward': 0.20494306087493896, 'reward_std': 0.0963810607790947, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1939438134431839, 'sampling/sampling_logp_difference/max': 1.908888816833496, 'sampling/importance_sampling_ratio/min': 0.14824502170085907, 'sampling/importance_sampling_ratio/mean': 1.037878394126892, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2233941107988358, 'clip_ratio/low_mean': 0.08701690286397934, 'clip_ratio/low_min': 0.08701690286397934, 'clip_ratio/high_mean': 0.0969635471701622, 'clip_ratio/high_max': 0.0969635471701622, 'clip_ratio/region_mean': 0.18398045003414154, 'reward_total_mean': 0.20494306087493896, 'reward_meter_mean': 0.7604238390922546, 'reward_meter_std': 0.27199587225914, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9908215999603271, 'reward_repeat_penalty_std': 0.009297901764512062, 'reward_near_duplicate_penalty_mean': 0.9939049482345581, 'reward_near_duplicate_penalty_std': 0.0034192020539194345, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9968982934951782, 'reward_distinct_2_std': 0.008773035369813442, 'reward_judge_quality_mean': 0.3375000059604645, 'reward_judge_quality_std': 0.08345229178667068, 'reward_total_composite_mean': 0.20494306087493896, 'reward_total_composite_std': 0.0963810607790947, 'epoch': 0.0}

  2%|▏         | 78/3300 [16:08<9:51:21, 11.01s/it]INFO 04-12 11:02:20 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 79/3300 [16:20<10:06:49, 11.30s/it]2026-04-12 11:02:32,390 | INFO | train_grpo_train | metrics_logged mode=train step=79

                                                    
{'loss': -0.0354, 'grad_norm': 6.8120436668396, 'learning_rate': 9.763636363636365e-06, 'num_tokens': 176653.0, 'completions/mean_length': 196.0, 'completions/min_length': 103.0, 'completions/max_length': 254.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 196.0, 'completions/min_terminated_length': 103.0, 'completions/max_terminated_length': 254.0, 'rewards/meter/mean': 0.6108548045158386, 'rewards/meter/std': 0.36881759762763977, 'rewards/count_adherence/mean': 0.8541666269302368, 'rewards/count_adherence/std': 0.0589255727827549, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9920930862426758, 'rewards/lexical_plausibility/std': 0.022364171221852303, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.11649646610021591, 'rewards/repeat_penalty/mean': 0.9866504073143005, 'rewards/repeat_penalty/std': 0.016313185915350914, 'rewards/near_duplicate_penalty/mean': 0.9969182014465332, 'rewards/near_duplicate_penalty/std': 0.0017705624923110008, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.989695131778717, 'rewards/distinct_2/std': 0.015878980979323387, 'rewards/total_composite/mean': 0.16621604561805725, 'rewards/total_composite/std': 0.1380038559436798, 'reward': 0.16621604561805725, 'reward_std': 0.13800382614135742, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2137884944677353, 'sampling/sampling_logp_difference/max': 1.9605722427368164, 'sampling/importance_sampling_ratio/min': 0.1407778412103653, 'sampling/importance_sampling_ratio/mean': 1.0485703945159912, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.92401659488678, 'clip_ratio/low_mean': 0.10013115033507347, 'clip_ratio/low_min': 0.10013115033507347, 'clip_ratio/high_mean': 0.0744392741471529, 'clip_ratio/high_max': 0.0744392741471529, 'clip_ratio/region_mean': 0.17457042448222637, 'reward_total_mean': 0.16621604561805725, 'reward_meter_mean': 0.6108548045158386, 'reward_meter_std': 0.36881759762763977, 'reward_count_adherence_mean': 0.8541666269302368, 'reward_count_adherence_std': 0.0589255727827549, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9920930862426758, 'reward_lexical_plausibility_std': 0.022364171221852303, 'reward_repeat_penalty_mean': 0.9866504073143005, 'reward_repeat_penalty_std': 0.016313185915350914, 'reward_near_duplicate_penalty_mean': 0.9969182014465332, 'reward_near_duplicate_penalty_std': 0.0017705624923110008, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.989695131778717, 'reward_distinct_2_std': 0.015878980979323387, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.11649646610021591, 'reward_total_composite_mean': 0.16621604561805725, 'reward_total_composite_std': 0.1380038559436798, 'epoch': 0.0}

  2%|▏         | 79/3300 [16:20<10:06:49, 11.30s/it]INFO 04-12 11:02:32 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 80/3300 [16:29<9:35:12, 10.72s/it] 2026-04-12 11:02:41,742 | INFO | train_grpo_train | metrics_logged mode=train step=80

                                                   
{'loss': 0.0156, 'grad_norm': 15.379478454589844, 'learning_rate': 9.760606060606062e-06, 'num_tokens': 178337.0, 'completions/mean_length': 45.5, 'completions/min_length': 34.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.5, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.6371452212333679, 'rewards/meter/std': 0.42659109830856323, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5287500023841858, 'rewards/judge_quality/std': 0.18333713710308075, 'rewards/repeat_penalty/mean': 0.9953550696372986, 'rewards/repeat_penalty/std': 0.006389432121068239, 'rewards/near_duplicate_penalty/mean': 0.9953550696372986, 'rewards/near_duplicate_penalty/std': 0.006389432121068239, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3356579542160034, 'rewards/total_composite/std': 0.2821563184261322, 'reward': 0.3356579542160034, 'reward_std': 0.2821563184261322, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21456630527973175, 'sampling/sampling_logp_difference/max': 1.4055638313293457, 'sampling/importance_sampling_ratio/min': 0.24522875249385834, 'sampling/importance_sampling_ratio/mean': 1.0313081741333008, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.3910775184631348, 'clip_ratio/low_mean': 0.0916936518624425, 'clip_ratio/low_min': 0.0916936518624425, 'clip_ratio/high_mean': 0.1300747711211443, 'clip_ratio/high_max': 0.1300747711211443, 'clip_ratio/region_mean': 0.2217684229835868, 'reward_total_mean': 0.3356579542160034, 'reward_meter_mean': 0.6371452212333679, 'reward_meter_std': 0.42659109830856323, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9953550696372986, 'reward_repeat_penalty_std': 0.006389432121068239, 'reward_near_duplicate_penalty_mean': 0.9953550696372986, 'reward_near_duplicate_penalty_std': 0.006389432121068239, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5287500023841858, 'reward_judge_quality_std': 0.18333713710308075, 'reward_total_composite_mean': 0.3356579542160034, 'reward_total_composite_std': 0.2821563184261322, 'epoch': 0.0}

  2%|▏         | 80/3300 [16:29<9:35:12, 10.72s/it]INFO 04-12 11:02:41 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 81/3300 [16:39<9:15:18, 10.35s/it]2026-04-12 11:02:51,235 | INFO | train_grpo_train | metrics_logged mode=train step=81

                                                   
{'loss': 0.0635, 'grad_norm': 13.21326732635498, 'learning_rate': 9.757575757575758e-06, 'num_tokens': 180108.0, 'completions/mean_length': 48.375, 'completions/min_length': 35.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.375, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.3304923474788666, 'rewards/meter/std': 0.3466903567314148, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9575716257095337, 'rewards/lexical_plausibility/std': 0.0786299780011177, 'rewards/judge_quality/mean': 0.5049999952316284, 'rewards/judge_quality/std': 0.2584569752216339, 'rewards/repeat_penalty/mean': 0.9907418489456177, 'rewards/repeat_penalty/std': 0.01250008586794138, 'rewards/near_duplicate_penalty/mean': 0.9907418489456177, 'rewards/near_duplicate_penalty/std': 0.01250008586794138, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1738933026790619, 'rewards/total_composite/std': 0.22043001651763916, 'reward': 0.1738933026790619, 'reward_std': 0.22043000161647797, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2102905660867691, 'sampling/sampling_logp_difference/max': 1.1979718208312988, 'sampling/importance_sampling_ratio/min': 0.301805704832077, 'sampling/importance_sampling_ratio/mean': 1.054197907447815, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.183119833469391, 'clip_ratio/low_mean': 0.13772014062851667, 'clip_ratio/low_min': 0.13772014062851667, 'clip_ratio/high_mean': 0.07689860463142395, 'clip_ratio/high_max': 0.07689860463142395, 'clip_ratio/region_mean': 0.21461874525994062, 'reward_total_mean': 0.1738933026790619, 'reward_meter_mean': 0.3304923474788666, 'reward_meter_std': 0.3466903567314148, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9575716257095337, 'reward_lexical_plausibility_std': 0.0786299780011177, 'reward_repeat_penalty_mean': 0.9907418489456177, 'reward_repeat_penalty_std': 0.01250008586794138, 'reward_near_duplicate_penalty_mean': 0.9907418489456177, 'reward_near_duplicate_penalty_std': 0.01250008586794138, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5049999952316284, 'reward_judge_quality_std': 0.2584569752216339, 'reward_total_composite_mean': 0.1738933026790619, 'reward_total_composite_std': 0.22043001651763916, 'epoch': 0.0}

  2%|▏         | 81/3300 [16:39<9:15:18, 10.35s/it]INFO 04-12 11:02:51 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 82/3300 [16:48<9:00:58, 10.09s/it]2026-04-12 11:03:00,706 | INFO | train_grpo_train | metrics_logged mode=train step=82

                                                   
{'loss': 0.0446, 'grad_norm': 20.60401153564453, 'learning_rate': 9.754545454545455e-06, 'num_tokens': 181911.0, 'completions/mean_length': 62.375, 'completions/min_length': 52.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.375, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.5103211998939514, 'rewards/meter/std': 0.457789808511734, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.10690449178218842, 'rewards/repeat_penalty/mean': 0.972140371799469, 'rewards/repeat_penalty/std': 0.03721826151013374, 'rewards/near_duplicate_penalty/mean': 0.9814019203186035, 'rewards/near_duplicate_penalty/std': 0.01771257445216179, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.990384578704834, 'rewards/distinct_2/std': 0.027196412906050682, 'rewards/total_composite/mean': 0.16879403591156006, 'rewards/total_composite/std': 0.17233037948608398, 'reward': 0.16879403591156006, 'reward_std': 0.17233037948608398, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20648972690105438, 'sampling/sampling_logp_difference/max': 2.3103561401367188, 'sampling/importance_sampling_ratio/min': 0.09922590851783752, 'sampling/importance_sampling_ratio/mean': 0.9993157386779785, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8028803467750549, 'clip_ratio/low_mean': 0.10225484520196915, 'clip_ratio/low_min': 0.10225484520196915, 'clip_ratio/high_mean': 0.04357881844043732, 'clip_ratio/high_max': 0.04357881844043732, 'clip_ratio/region_mean': 0.14583366364240646, 'reward_total_mean': 0.16879403591156006, 'reward_meter_mean': 0.5103211998939514, 'reward_meter_std': 0.457789808511734, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.972140371799469, 'reward_repeat_penalty_std': 0.03721826151013374, 'reward_near_duplicate_penalty_mean': 0.9814019203186035, 'reward_near_duplicate_penalty_std': 0.01771257445216179, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.990384578704834, 'reward_distinct_2_std': 0.027196412906050682, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.10690449178218842, 'reward_total_composite_mean': 0.16879403591156006, 'reward_total_composite_std': 0.17233037948608398, 'epoch': 0.0}

  2%|▏         | 82/3300 [16:48<9:00:58, 10.09s/it]INFO 04-12 11:03:00 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 83/3300 [16:58<8:53:10,  9.94s/it]2026-04-12 11:03:10,318 | INFO | train_grpo_train | metrics_logged mode=train step=83

                                                   
{'loss': -0.0125, 'grad_norm': 11.30092716217041, 'learning_rate': 9.751515151515152e-06, 'num_tokens': 183548.0, 'completions/mean_length': 51.625, 'completions/min_length': 35.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.625, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.5117467641830444, 'rewards/meter/std': 0.42367249727249146, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9543439149856567, 'rewards/lexical_plausibility/std': 0.10052025318145752, 'rewards/judge_quality/mean': 0.5887500047683716, 'rewards/judge_quality/std': 0.27740830183029175, 'rewards/repeat_penalty/mean': 0.9962121248245239, 'rewards/repeat_penalty/std': 0.007013781927525997, 'rewards/near_duplicate_penalty/mean': 0.9962121248245239, 'rewards/near_duplicate_penalty/std': 0.007013781927525997, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2862284779548645, 'rewards/total_composite/std': 0.2962500751018524, 'reward': 0.2862284779548645, 'reward_std': 0.2962500751018524, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18761391937732697, 'sampling/sampling_logp_difference/max': 1.7637081146240234, 'sampling/importance_sampling_ratio/min': 0.17140808701515198, 'sampling/importance_sampling_ratio/mean': 1.0273298025131226, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.882716327905655, 'clip_ratio/low_mean': 0.07597680296748877, 'clip_ratio/low_min': 0.07597680296748877, 'clip_ratio/high_mean': 0.1147812120616436, 'clip_ratio/high_max': 0.1147812120616436, 'clip_ratio/region_mean': 0.19075801502913237, 'reward_total_mean': 0.2862284779548645, 'reward_meter_mean': 0.5117467641830444, 'reward_meter_std': 0.42367249727249146, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9543439149856567, 'reward_lexical_plausibility_std': 0.10052025318145752, 'reward_repeat_penalty_mean': 0.9962121248245239, 'reward_repeat_penalty_std': 0.007013781927525997, 'reward_near_duplicate_penalty_mean': 0.9962121248245239, 'reward_near_duplicate_penalty_std': 0.007013781927525997, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5887500047683716, 'reward_judge_quality_std': 0.27740830183029175, 'reward_total_composite_mean': 0.2862284779548645, 'reward_total_composite_std': 0.2962500751018524, 'epoch': 0.0}

  3%|▎         | 83/3300 [16:58<8:53:10,  9.94s/it]INFO 04-12 11:03:10 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 84/3300 [17:07<8:36:04,  9.63s/it]2026-04-12 11:03:19,210 | INFO | train_grpo_train | metrics_logged mode=train step=84

                                                   
{'loss': 0.1781, 'grad_norm': 15.115890502929688, 'learning_rate': 9.74848484848485e-06, 'num_tokens': 185027.0, 'completions/mean_length': 43.875, 'completions/min_length': 34.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.875, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.8157832622528076, 'rewards/meter/std': 0.28191104531288147, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5549999475479126, 'rewards/judge_quality/std': 0.24599653482437134, 'rewards/repeat_penalty/mean': 0.9908944368362427, 'rewards/repeat_penalty/std': 0.01328963227570057, 'rewards/near_duplicate_penalty/mean': 0.9908944368362427, 'rewards/near_duplicate_penalty/std': 0.01328963227570057, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.47203630208969116, 'rewards/total_composite/std': 0.2735925018787384, 'reward': 0.47203630208969116, 'reward_std': 0.2735925316810608, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20224647223949432, 'sampling/sampling_logp_difference/max': 1.2447290420532227, 'sampling/importance_sampling_ratio/min': 0.28801894187927246, 'sampling/importance_sampling_ratio/mean': 1.0262371301651, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5166660696268082, 'clip_ratio/low_mean': 0.09913571737706661, 'clip_ratio/low_min': 0.09913571737706661, 'clip_ratio/high_mean': 0.05674962420016527, 'clip_ratio/high_max': 0.05674962420016527, 'clip_ratio/region_mean': 0.15588534157723188, 'reward_total_mean': 0.47203630208969116, 'reward_meter_mean': 0.8157832622528076, 'reward_meter_std': 0.28191104531288147, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9908944368362427, 'reward_repeat_penalty_std': 0.01328963227570057, 'reward_near_duplicate_penalty_mean': 0.9908944368362427, 'reward_near_duplicate_penalty_std': 0.01328963227570057, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5549999475479126, 'reward_judge_quality_std': 0.24599653482437134, 'reward_total_composite_mean': 0.47203630208969116, 'reward_total_composite_std': 0.2735925018787384, 'epoch': 0.0}

  3%|▎         | 84/3300 [17:07<8:36:04,  9.63s/it]INFO 04-12 11:03:19 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 85/3300 [17:15<8:17:20,  9.28s/it]2026-04-12 11:03:27,682 | INFO | train_grpo_train | metrics_logged mode=train step=85

                                                   
{'loss': 0.0799, 'grad_norm': 15.081714630126953, 'learning_rate': 9.745454545454547e-06, 'num_tokens': 186613.0, 'completions/mean_length': 47.25, 'completions/min_length': 37.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.25, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.7033107280731201, 'rewards/meter/std': 0.40071213245391846, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.5049999952316284, 'rewards/judge_quality/std': 0.2584569752216339, 'rewards/repeat_penalty/mean': 0.9889481663703918, 'rewards/repeat_penalty/std': 0.009406981989741325, 'rewards/near_duplicate_penalty/mean': 0.9889481663703918, 'rewards/near_duplicate_penalty/std': 0.009406981989741325, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.39188358187675476, 'rewards/total_composite/std': 0.34003448486328125, 'reward': 0.39188358187675476, 'reward_std': 0.34003445506095886, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17086577415466309, 'sampling/sampling_logp_difference/max': 2.290557384490967, 'sampling/importance_sampling_ratio/min': 0.10121003538370132, 'sampling/importance_sampling_ratio/mean': 1.0280930995941162, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5397960841655731, 'clip_ratio/low_mean': 0.10041468869894743, 'clip_ratio/low_min': 0.10041468869894743, 'clip_ratio/high_mean': 0.036472697742283344, 'clip_ratio/high_max': 0.036472697742283344, 'clip_ratio/region_mean': 0.13688738644123077, 'reward_total_mean': 0.39188358187675476, 'reward_meter_mean': 0.7033107280731201, 'reward_meter_std': 0.40071213245391846, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.9889481663703918, 'reward_repeat_penalty_std': 0.009406981989741325, 'reward_near_duplicate_penalty_mean': 0.9889481663703918, 'reward_near_duplicate_penalty_std': 0.009406981989741325, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5049999952316284, 'reward_judge_quality_std': 0.2584569752216339, 'reward_total_composite_mean': 0.39188358187675476, 'reward_total_composite_std': 0.34003448486328125, 'epoch': 0.0}

  3%|▎         | 85/3300 [17:15<8:17:20,  9.28s/it]INFO 04-12 11:03:27 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 86/3300 [17:25<8:26:23,  9.45s/it]2026-04-12 11:03:37,535 | INFO | train_grpo_train | metrics_logged mode=train step=86

                                                   
{'loss': 0.071, 'grad_norm': 11.537137031555176, 'learning_rate': 9.742424242424244e-06, 'num_tokens': 188462.0, 'completions/mean_length': 62.125, 'completions/min_length': 38.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.125, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.9878455996513367, 'rewards/meter/std': 0.01681375503540039, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.4087499976158142, 'rewards/judge_quality/std': 0.22273865342140198, 'rewards/repeat_penalty/mean': 0.9923967123031616, 'rewards/repeat_penalty/std': 0.013055448420345783, 'rewards/near_duplicate_penalty/mean': 0.9923967123031616, 'rewards/near_duplicate_penalty/std': 0.013055448420345783, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.40400809049606323, 'rewards/total_composite/std': 0.2218370884656906, 'reward': 0.40400809049606323, 'reward_std': 0.2218370884656906, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19538550078868866, 'sampling/sampling_logp_difference/max': 1.8238182067871094, 'sampling/importance_sampling_ratio/min': 0.16140829026699066, 'sampling/importance_sampling_ratio/mean': 1.0408320426940918, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1176094114780426, 'clip_ratio/low_mean': 0.124702257104218, 'clip_ratio/low_min': 0.124702257104218, 'clip_ratio/high_mean': 0.03133971244096756, 'clip_ratio/high_max': 0.03133971244096756, 'clip_ratio/region_mean': 0.15604196954518557, 'reward_total_mean': 0.40400809049606323, 'reward_meter_mean': 0.9878455996513367, 'reward_meter_std': 0.01681375503540039, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.9923967123031616, 'reward_repeat_penalty_std': 0.013055448420345783, 'reward_near_duplicate_penalty_mean': 0.9923967123031616, 'reward_near_duplicate_penalty_std': 0.013055448420345783, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4087499976158142, 'reward_judge_quality_std': 0.22273865342140198, 'reward_total_composite_mean': 0.40400809049606323, 'reward_total_composite_std': 0.2218370884656906, 'epoch': 0.0}

  3%|▎         | 86/3300 [17:25<8:26:23,  9.45s/it]INFO 04-12 11:03:37 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 87/3300 [17:34<8:25:33,  9.44s/it]2026-04-12 11:03:46,947 | INFO | train_grpo_train | metrics_logged mode=train step=87

                                                   
{'loss': -0.0532, 'grad_norm': 12.620993614196777, 'learning_rate': 9.739393939393941e-06, 'num_tokens': 190139.0, 'completions/mean_length': 53.625, 'completions/min_length': 38.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.625, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.750113844871521, 'rewards/meter/std': 0.3385850489139557, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.9992472529411316, 'rewards/repeat_penalty/std': 0.0014177416451275349, 'rewards/near_duplicate_penalty/mean': 0.9992472529411316, 'rewards/near_duplicate_penalty/std': 0.0014177416451275349, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.23165175318717957, 'rewards/total_composite/std': 0.15431727468967438, 'reward': 0.23165175318717957, 'reward_std': 0.15431728959083557, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22738726437091827, 'sampling/sampling_logp_difference/max': 1.3244447708129883, 'sampling/importance_sampling_ratio/min': 0.26595059037208557, 'sampling/importance_sampling_ratio/mean': 1.0289357900619507, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.525059252977371, 'clip_ratio/low_mean': 0.13426264375448227, 'clip_ratio/low_min': 0.13426264375448227, 'clip_ratio/high_mean': 0.08101851865649223, 'clip_ratio/high_max': 0.08101851865649223, 'clip_ratio/region_mean': 0.2152811624109745, 'reward_total_mean': 0.23165175318717957, 'reward_meter_mean': 0.750113844871521, 'reward_meter_std': 0.3385850489139557, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9992472529411316, 'reward_repeat_penalty_std': 0.0014177416451275349, 'reward_near_duplicate_penalty_mean': 0.9992472529411316, 'reward_near_duplicate_penalty_std': 0.0014177416451275349, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.23165175318717957, 'reward_total_composite_std': 0.15431727468967438, 'epoch': 0.0}

  3%|▎         | 87/3300 [17:34<8:25:33,  9.44s/it]INFO 04-12 11:03:47 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 88/3300 [17:44<8:34:02,  9.60s/it]2026-04-12 11:03:56,926 | INFO | train_grpo_train | metrics_logged mode=train step=88

                                                   
{'loss': 0.0411, 'grad_norm': 10.12899398803711, 'learning_rate': 9.736363636363637e-06, 'num_tokens': 192516.0, 'completions/mean_length': 104.125, 'completions/min_length': 81.0, 'completions/max_length': 135.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 104.125, 'completions/min_terminated_length': 81.0, 'completions/max_terminated_length': 135.0, 'rewards/meter/mean': 0.5040664076805115, 'rewards/meter/std': 0.3933837115764618, '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/lexical_plausibility/mean': 0.9983552694320679, 'rewards/lexical_plausibility/std': 0.004652021918445826, 'rewards/judge_quality/mean': 0.4124999940395355, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/repeat_penalty/mean': 0.9975301623344421, 'rewards/repeat_penalty/std': 0.0026607722975313663, 'rewards/near_duplicate_penalty/mean': 0.9975301623344421, 'rewards/near_duplicate_penalty/std': 0.0026607722975313663, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.21353963017463684, 'rewards/total_composite/std': 0.18697339296340942, 'reward': 0.21353963017463684, 'reward_std': 0.18697339296340942, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20189937949180603, 'sampling/sampling_logp_difference/max': 1.740591049194336, 'sampling/importance_sampling_ratio/min': 0.17541669309139252, 'sampling/importance_sampling_ratio/mean': 1.040609359741211, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.331147611141205, 'clip_ratio/low_mean': 0.10875261202454567, 'clip_ratio/low_min': 0.10875261202454567, 'clip_ratio/high_mean': 0.08587215840816498, 'clip_ratio/high_max': 0.08587215840816498, 'clip_ratio/region_mean': 0.19462477043271065, 'reward_total_mean': 0.21353963017463684, 'reward_meter_mean': 0.5040664076805115, 'reward_meter_std': 0.3933837115764618, '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_lexical_plausibility_mean': 0.9983552694320679, 'reward_lexical_plausibility_std': 0.004652021918445826, 'reward_repeat_penalty_mean': 0.9975301623344421, 'reward_repeat_penalty_std': 0.0026607722975313663, 'reward_near_duplicate_penalty_mean': 0.9975301623344421, 'reward_near_duplicate_penalty_std': 0.0026607722975313663, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4124999940395355, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.21353963017463684, 'reward_total_composite_std': 0.18697339296340942, 'epoch': 0.0}

  3%|▎         | 88/3300 [17:44<8:34:02,  9.60s/it]INFO 04-12 11:03:57 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 89/3300 [17:54<8:29:02,  9.51s/it]2026-04-12 11:04:06,246 | INFO | train_grpo_train | metrics_logged mode=train step=89

                                                   
{'loss': 0.1083, 'grad_norm': 16.61884117126465, 'learning_rate': 9.733333333333334e-06, 'num_tokens': 194080.0, 'completions/mean_length': 45.5, 'completions/min_length': 30.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.5, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.45840001106262207, 'rewards/meter/std': 0.38367897272109985, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6462500095367432, 'rewards/judge_quality/std': 0.3168567717075348, 'rewards/repeat_penalty/mean': 0.9884709119796753, 'rewards/repeat_penalty/std': 0.016373859718441963, 'rewards/near_duplicate_penalty/mean': 0.9884709119796753, 'rewards/near_duplicate_penalty/std': 0.016373859718441963, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.40065518021583557, 'rewards/total_composite/std': 0.3918873071670532, 'reward': 0.40065518021583557, 'reward_std': 0.3918873071670532, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18994733691215515, 'sampling/sampling_logp_difference/max': 2.3873844146728516, 'sampling/importance_sampling_ratio/min': 0.09186965972185135, 'sampling/importance_sampling_ratio/mean': 1.021990180015564, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4212532937526703, 'clip_ratio/low_mean': 0.09404137637466192, 'clip_ratio/low_min': 0.09404137637466192, 'clip_ratio/high_mean': 0.07141768280416727, 'clip_ratio/high_max': 0.07141768280416727, 'clip_ratio/region_mean': 0.1654590591788292, 'reward_total_mean': 0.40065518021583557, 'reward_meter_mean': 0.45840001106262207, 'reward_meter_std': 0.38367897272109985, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9884709119796753, 'reward_repeat_penalty_std': 0.016373859718441963, 'reward_near_duplicate_penalty_mean': 0.9884709119796753, 'reward_near_duplicate_penalty_std': 0.016373859718441963, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6462500095367432, 'reward_judge_quality_std': 0.3168567717075348, 'reward_total_composite_mean': 0.40065518021583557, 'reward_total_composite_std': 0.3918873071670532, 'epoch': 0.0}

  3%|▎         | 89/3300 [17:54<8:29:02,  9.51s/it]INFO 04-12 11:04:06 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 90/3300 [18:12<10:51:29, 12.18s/it]2026-04-12 11:04:24,626 | INFO | train_grpo_train | metrics_logged mode=train step=90

                                                    
{'loss': 0.0224, 'grad_norm': 7.058398723602295, 'learning_rate': 9.730303030303031e-06, 'num_tokens': 197197.0, 'completions/mean_length': 194.625, 'completions/min_length': 137.0, 'completions/max_length': 298.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 194.625, 'completions/min_terminated_length': 137.0, 'completions/max_terminated_length': 298.0, 'rewards/meter/mean': 0.6234270334243774, 'rewards/meter/std': 0.26951634883880615, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.06681530922651291, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.36249998211860657, 'rewards/judge_quality/std': 0.11259915679693222, 'rewards/repeat_penalty/mean': 0.9690656065940857, 'rewards/repeat_penalty/std': 0.022364940494298935, 'rewards/near_duplicate_penalty/mean': 0.9883231520652771, 'rewards/near_duplicate_penalty/std': 0.006093398667871952, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.980473518371582, 'rewards/distinct_2/std': 0.019685424864292145, 'rewards/total_composite/mean': 0.2098064124584198, 'rewards/total_composite/std': 0.13666240870952606, 'reward': 0.2098064124584198, 'reward_std': 0.13666240870952606, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22589033842086792, 'sampling/sampling_logp_difference/max': 2.057332992553711, 'sampling/importance_sampling_ratio/min': 0.12779435515403748, 'sampling/importance_sampling_ratio/mean': 1.0478851795196533, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.973706364631653, 'clip_ratio/low_mean': 0.09569885954260826, 'clip_ratio/low_min': 0.09569885954260826, 'clip_ratio/high_mean': 0.09893776290118694, 'clip_ratio/high_max': 0.09893776290118694, 'clip_ratio/region_mean': 0.1946366224437952, 'reward_total_mean': 0.2098064124584198, 'reward_meter_mean': 0.6234270334243774, 'reward_meter_std': 0.26951634883880615, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.06681530922651291, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9690656065940857, 'reward_repeat_penalty_std': 0.022364940494298935, 'reward_near_duplicate_penalty_mean': 0.9883231520652771, 'reward_near_duplicate_penalty_std': 0.006093398667871952, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.980473518371582, 'reward_distinct_2_std': 0.019685424864292145, 'reward_judge_quality_mean': 0.36249998211860657, 'reward_judge_quality_std': 0.11259915679693222, 'reward_total_composite_mean': 0.2098064124584198, 'reward_total_composite_std': 0.13666240870952606, 'epoch': 0.0}

  3%|▎         | 90/3300 [18:12<10:51:29, 12.18s/it]INFO 04-12 11:04:24 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 91/3300 [18:25<11:09:28, 12.52s/it]2026-04-12 11:04:37,936 | INFO | train_grpo_train | metrics_logged mode=train step=91

                                                    
{'loss': 0.1641, 'grad_norm': 6.941081523895264, 'learning_rate': 9.727272727272728e-06, 'num_tokens': 200683.0, 'completions/mean_length': 237.75, 'completions/min_length': 168.0, 'completions/max_length': 355.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 237.75, 'completions/min_terminated_length': 168.0, 'completions/max_terminated_length': 355.0, 'rewards/meter/mean': 0.502982497215271, 'rewards/meter/std': 0.38829395174980164, 'rewards/count_adherence/mean': 0.8888888955116272, 'rewards/count_adherence/std': 0.059391383081674576, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9970703125, 'rewards/lexical_plausibility/std': 0.008286407217383385, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.9909600019454956, 'rewards/repeat_penalty/std': 0.006515172775834799, 'rewards/near_duplicate_penalty/mean': 0.9931068420410156, 'rewards/near_duplicate_penalty/std': 0.0028236538637429476, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9978392124176025, 'rewards/distinct_2/std': 0.006111545953899622, 'rewards/total_composite/mean': 0.14956414699554443, 'rewards/total_composite/std': 0.12625251710414886, 'reward': 0.14956414699554443, 'reward_std': 0.12625251710414886, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21524831652641296, 'sampling/sampling_logp_difference/max': 1.6496601104736328, 'sampling/importance_sampling_ratio/min': 0.1921152025461197, 'sampling/importance_sampling_ratio/mean': 1.0547246932983398, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.1639786660671234, 'clip_ratio/low_mean': 0.12241872772574425, 'clip_ratio/low_min': 0.12241872772574425, 'clip_ratio/high_mean': 0.07882829755544662, 'clip_ratio/high_max': 0.07882829755544662, 'clip_ratio/region_mean': 0.20124702528119087, 'reward_total_mean': 0.14956414699554443, 'reward_meter_mean': 0.502982497215271, 'reward_meter_std': 0.38829395174980164, 'reward_count_adherence_mean': 0.8888888955116272, 'reward_count_adherence_std': 0.059391383081674576, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9970703125, 'reward_lexical_plausibility_std': 0.008286407217383385, 'reward_repeat_penalty_mean': 0.9909600019454956, 'reward_repeat_penalty_std': 0.006515172775834799, 'reward_near_duplicate_penalty_mean': 0.9931068420410156, 'reward_near_duplicate_penalty_std': 0.0028236538637429476, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9978392124176025, 'reward_distinct_2_std': 0.006111545953899622, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.14956414699554443, 'reward_total_composite_std': 0.12625251710414886, 'epoch': 0.0}

  3%|▎         | 91/3300 [18:25<11:09:28, 12.52s/it]INFO 04-12 11:04:38 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 92/3300 [18:38<11:04:10, 12.42s/it]2026-04-12 11:04:50,137 | INFO | train_grpo_train | metrics_logged mode=train step=92

                                                    
{'loss': 0.2234, 'grad_norm': 6.573692321777344, 'learning_rate': 9.724242424242426e-06, 'num_tokens': 204357.0, 'completions/mean_length': 218.25, 'completions/min_length': 114.0, 'completions/max_length': 293.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 218.25, 'completions/min_terminated_length': 114.0, 'completions/max_terminated_length': 293.0, 'rewards/meter/mean': 0.5156489610671997, 'rewards/meter/std': 0.40056365728378296, 'rewards/count_adherence/mean': 0.9107142686843872, 'rewards/count_adherence/std': 0.07393559068441391, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9735461473464966, 'rewards/lexical_plausibility/std': 0.049191076308488846, 'rewards/judge_quality/mean': 0.1875, 'rewards/judge_quality/std': 0.07440237700939178, 'rewards/repeat_penalty/mean': 0.9880360960960388, 'rewards/repeat_penalty/std': 0.011319074779748917, 'rewards/near_duplicate_penalty/mean': 0.9935440421104431, 'rewards/near_duplicate_penalty/std': 0.0037526292726397514, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9944518804550171, 'rewards/distinct_2/std': 0.010290447622537613, 'rewards/total_composite/mean': 0.09464669227600098, 'rewards/total_composite/std': 0.0865207239985466, 'reward': 0.09464669227600098, 'reward_std': 0.0865207239985466, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2075955718755722, 'sampling/sampling_logp_difference/max': 3.3988921642303467, 'sampling/importance_sampling_ratio/min': 0.03341026231646538, 'sampling/importance_sampling_ratio/mean': 1.05128014087677, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.64172226190567, 'clip_ratio/low_mean': 0.0901694968342781, 'clip_ratio/low_min': 0.0901694968342781, 'clip_ratio/high_mean': 0.10005187802016735, 'clip_ratio/high_max': 0.10005187802016735, 'clip_ratio/region_mean': 0.19022137485444546, 'reward_total_mean': 0.09464669227600098, 'reward_meter_mean': 0.5156489610671997, 'reward_meter_std': 0.40056365728378296, 'reward_count_adherence_mean': 0.9107142686843872, 'reward_count_adherence_std': 0.07393559068441391, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9735461473464966, 'reward_lexical_plausibility_std': 0.049191076308488846, 'reward_repeat_penalty_mean': 0.9880360960960388, 'reward_repeat_penalty_std': 0.011319074779748917, 'reward_near_duplicate_penalty_mean': 0.9935440421104431, 'reward_near_duplicate_penalty_std': 0.0037526292726397514, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9944518804550171, 'reward_distinct_2_std': 0.010290447622537613, 'reward_judge_quality_mean': 0.1875, 'reward_judge_quality_std': 0.07440237700939178, 'reward_total_composite_mean': 0.09464669227600098, 'reward_total_composite_std': 0.0865207239985466, 'epoch': 0.0}

  3%|▎         | 92/3300 [18:38<11:04:10, 12.42s/it]INFO 04-12 11:04:50 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 93/3300 [18:55<12:28:34, 14.01s/it]2026-04-12 11:05:08,096 | INFO | train_grpo_train | metrics_logged mode=train step=93

                                                    
{'loss': 0.0032, 'grad_norm': 8.9693603515625, 'learning_rate': 9.721212121212123e-06, 'num_tokens': 206682.0, 'completions/mean_length': 109.625, 'completions/min_length': 74.0, 'completions/max_length': 140.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 109.625, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 140.0, 'rewards/meter/mean': 0.8595813512802124, 'rewards/meter/std': 0.1380358338356018, '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/lexical_plausibility/mean': 0.9942992329597473, 'rewards/lexical_plausibility/std': 0.010885572992265224, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.9930427670478821, 'rewards/repeat_penalty/std': 0.007280769292265177, 'rewards/near_duplicate_penalty/mean': 0.9930427670478821, 'rewards/near_duplicate_penalty/std': 0.007280769292265177, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.19612924754619598, 'rewards/total_composite/std': 0.13066346943378448, 'reward': 0.19612924754619598, 'reward_std': 0.13066346943378448, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22369436919689178, 'sampling/sampling_logp_difference/max': 1.2494277954101562, 'sampling/importance_sampling_ratio/min': 0.2866687774658203, 'sampling/importance_sampling_ratio/mean': 1.0495545864105225, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.9556396305561066, 'clip_ratio/low_mean': 0.08056333288550377, 'clip_ratio/low_min': 0.08056333288550377, 'clip_ratio/high_mean': 0.11009310558438301, 'clip_ratio/high_max': 0.11009310558438301, 'clip_ratio/region_mean': 0.19065643846988678, 'reward_total_mean': 0.19612924754619598, 'reward_meter_mean': 0.8595813512802124, 'reward_meter_std': 0.1380358338356018, '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_lexical_plausibility_mean': 0.9942992329597473, 'reward_lexical_plausibility_std': 0.010885572992265224, 'reward_repeat_penalty_mean': 0.9930427670478821, 'reward_repeat_penalty_std': 0.007280769292265177, 'reward_near_duplicate_penalty_mean': 0.9930427670478821, 'reward_near_duplicate_penalty_std': 0.007280769292265177, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.19612924754619598, 'reward_total_composite_std': 0.13066346943378448, 'epoch': 0.0}

  3%|▎         | 93/3300 [18:56<12:28:34, 14.01s/it]INFO 04-12 11:05:08 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 94/3300 [19:09<12:23:59, 13.92s/it]2026-04-12 11:05:21,570 | INFO | train_grpo_train | metrics_logged mode=train step=94

                                                    
{'loss': 0.0487, 'grad_norm': 15.944836616516113, 'learning_rate': 9.718181818181818e-06, 'num_tokens': 208244.0, 'completions/mean_length': 45.25, 'completions/min_length': 29.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.25, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.5198861956596375, 'rewards/meter/std': 0.4066314101219177, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9538085460662842, 'rewards/lexical_plausibility/std': 0.09894571453332901, 'rewards/judge_quality/mean': 0.45875000953674316, 'rewards/judge_quality/std': 0.192831352353096, 'rewards/repeat_penalty/mean': 0.9966422915458679, 'rewards/repeat_penalty/std': 0.007198222912847996, 'rewards/near_duplicate_penalty/mean': 0.9966422915458679, 'rewards/near_duplicate_penalty/std': 0.007198222912847996, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.26255714893341064, 'rewards/total_composite/std': 0.25084346532821655, 'reward': 0.26255714893341064, 'reward_std': 0.25084346532821655, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22381380200386047, 'sampling/sampling_logp_difference/max': 1.9445966482162476, 'sampling/importance_sampling_ratio/min': 0.14304490387439728, 'sampling/importance_sampling_ratio/mean': 1.0335097312927246, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.049190118908882, 'clip_ratio/low_mean': 0.0701514994725585, 'clip_ratio/low_min': 0.0701514994725585, 'clip_ratio/high_mean': 0.10071430541574955, 'clip_ratio/high_max': 0.10071430541574955, 'clip_ratio/region_mean': 0.17086580488830805, 'reward_total_mean': 0.26255714893341064, 'reward_meter_mean': 0.5198861956596375, 'reward_meter_std': 0.4066314101219177, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9538085460662842, 'reward_lexical_plausibility_std': 0.09894571453332901, 'reward_repeat_penalty_mean': 0.9966422915458679, 'reward_repeat_penalty_std': 0.007198222912847996, 'reward_near_duplicate_penalty_mean': 0.9966422915458679, 'reward_near_duplicate_penalty_std': 0.007198222912847996, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.45875000953674316, 'reward_judge_quality_std': 0.192831352353096, 'reward_total_composite_mean': 0.26255714893341064, 'reward_total_composite_std': 0.25084346532821655, 'epoch': 0.0}

  3%|▎         | 94/3300 [19:09<12:23:59, 13.92s/it]INFO 04-12 11:05:21 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 95/3300 [19:18<11:05:45, 12.46s/it]2026-04-12 11:05:30,626 | INFO | train_grpo_train | metrics_logged mode=train step=95

                                                    
{'loss': 0.0136, 'grad_norm': 15.552061080932617, 'learning_rate': 9.715151515151516e-06, 'num_tokens': 209937.0, 'completions/mean_length': 44.625, 'completions/min_length': 41.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.625, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.30856016278266907, 'rewards/meter/std': 0.3279573917388916, '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/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.42124998569488525, 'rewards/judge_quality/std': 0.22177450358867645, 'rewards/repeat_penalty/mean': 0.9786901473999023, 'rewards/repeat_penalty/std': 0.052769169211387634, 'rewards/near_duplicate_penalty/mean': 0.9886085987091064, 'rewards/near_duplicate_penalty/std': 0.02513744868338108, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9893162250518799, 'rewards/distinct_2/std': 0.03021823801100254, 'rewards/total_composite/mean': 0.15059463679790497, 'rewards/total_composite/std': 0.20923210680484772, 'reward': 0.15059463679790497, 'reward_std': 0.20923209190368652, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22380156815052032, 'sampling/sampling_logp_difference/max': 1.7575650215148926, 'sampling/importance_sampling_ratio/min': 0.1724643111228943, 'sampling/importance_sampling_ratio/mean': 1.0216649770736694, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.479477286338806, 'clip_ratio/low_mean': 0.17676972784101963, 'clip_ratio/low_min': 0.17676972784101963, 'clip_ratio/high_mean': 0.04884453862905502, 'clip_ratio/high_max': 0.04884453862905502, 'clip_ratio/region_mean': 0.22561426647007465, 'reward_total_mean': 0.15059463679790497, 'reward_meter_mean': 0.30856016278266907, 'reward_meter_std': 0.3279573917388916, '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_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.9786901473999023, 'reward_repeat_penalty_std': 0.052769169211387634, 'reward_near_duplicate_penalty_mean': 0.9886085987091064, 'reward_near_duplicate_penalty_std': 0.02513744868338108, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9893162250518799, 'reward_distinct_2_std': 0.03021823801100254, 'reward_judge_quality_mean': 0.42124998569488525, 'reward_judge_quality_std': 0.22177450358867645, 'reward_total_composite_mean': 0.15059463679790497, 'reward_total_composite_std': 0.20923210680484772, 'epoch': 0.0}

  3%|▎         | 95/3300 [19:18<11:05:45, 12.46s/it]INFO 04-12 11:05:30 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 96/3300 [19:28<10:26:36, 11.73s/it]2026-04-12 11:05:40,661 | INFO | train_grpo_train | metrics_logged mode=train step=96

                                                    
{'loss': 0.0391, 'grad_norm': 9.95379638671875, 'learning_rate': 9.712121212121213e-06, 'num_tokens': 212315.0, 'completions/mean_length': 120.25, 'completions/min_length': 87.0, 'completions/max_length': 135.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 120.25, 'completions/min_terminated_length': 87.0, 'completions/max_terminated_length': 135.0, 'rewards/meter/mean': 0.37057745456695557, 'rewards/meter/std': 0.25225669145584106, '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/lexical_plausibility/mean': 0.8845992684364319, 'rewards/lexical_plausibility/std': 0.1551179438829422, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.9951979517936707, 'rewards/repeat_penalty/std': 0.003440810600295663, 'rewards/near_duplicate_penalty/mean': 0.9951979517936707, 'rewards/near_duplicate_penalty/std': 0.003440810600295663, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.10789487510919571, 'rewards/total_composite/std': 0.07092024385929108, 'reward': 0.10789487510919571, 'reward_std': 0.07092024385929108, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21660031378269196, 'sampling/sampling_logp_difference/max': 1.989974021911621, 'sampling/importance_sampling_ratio/min': 0.13669897615909576, 'sampling/importance_sampling_ratio/mean': 1.0312293767929077, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.385300040245056, 'clip_ratio/low_mean': 0.0720515288412571, 'clip_ratio/low_min': 0.0720515288412571, 'clip_ratio/high_mean': 0.13101116940379143, 'clip_ratio/high_max': 0.13101116940379143, 'clip_ratio/region_mean': 0.20306269824504852, 'reward_total_mean': 0.10789487510919571, 'reward_meter_mean': 0.37057745456695557, 'reward_meter_std': 0.25225669145584106, '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_lexical_plausibility_mean': 0.8845992684364319, 'reward_lexical_plausibility_std': 0.1551179438829422, 'reward_repeat_penalty_mean': 0.9951979517936707, 'reward_repeat_penalty_std': 0.003440810600295663, 'reward_near_duplicate_penalty_mean': 0.9951979517936707, 'reward_near_duplicate_penalty_std': 0.003440810600295663, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.10789487510919571, 'reward_total_composite_std': 0.07092024385929108, 'epoch': 0.0}

  3%|▎         | 96/3300 [19:28<10:26:36, 11.73s/it]INFO 04-12 11:05:40 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 97/3300 [19:37<9:41:32, 10.89s/it] 2026-04-12 11:05:49,591 | INFO | train_grpo_train | metrics_logged mode=train step=97

                                                   
{'loss': 0.0193, 'grad_norm': 13.103679656982422, 'learning_rate': 9.70909090909091e-06, 'num_tokens': 214089.0, 'completions/mean_length': 52.75, 'completions/min_length': 44.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.75, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.509982705116272, 'rewards/meter/std': 0.4155988097190857, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6599999666213989, 'rewards/judge_quality/std': 0.2805097699165344, 'rewards/repeat_penalty/mean': 0.9952415227890015, 'rewards/repeat_penalty/std': 0.007189737632870674, 'rewards/near_duplicate_penalty/mean': 0.9952415227890015, 'rewards/near_duplicate_penalty/std': 0.007189737632870674, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.32618051767349243, 'rewards/total_composite/std': 0.31854259967803955, 'reward': 0.32618051767349243, 'reward_std': 0.31854259967803955, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18964336812496185, 'sampling/sampling_logp_difference/max': 1.374903678894043, 'sampling/importance_sampling_ratio/min': 0.25286394357681274, 'sampling/importance_sampling_ratio/mean': 1.00651216506958, 'sampling/importance_sampling_ratio/max': 1.9730236530303955, 'entropy': 1.791875198483467, 'clip_ratio/low_mean': 0.08039080258458853, 'clip_ratio/low_min': 0.08039080258458853, 'clip_ratio/high_mean': 0.12216202542185783, 'clip_ratio/high_max': 0.12216202542185783, 'clip_ratio/region_mean': 0.20255282800644636, 'reward_total_mean': 0.32618051767349243, 'reward_meter_mean': 0.509982705116272, 'reward_meter_std': 0.4155988097190857, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9952415227890015, 'reward_repeat_penalty_std': 0.007189737632870674, 'reward_near_duplicate_penalty_mean': 0.9952415227890015, 'reward_near_duplicate_penalty_std': 0.007189737632870674, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6599999666213989, 'reward_judge_quality_std': 0.2805097699165344, 'reward_total_composite_mean': 0.32618051767349243, 'reward_total_composite_std': 0.31854259967803955, 'epoch': 0.0}

  3%|▎         | 97/3300 [19:37<9:41:32, 10.89s/it]INFO 04-12 11:05:49 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 98/3300 [19:48<9:39:55, 10.87s/it]2026-04-12 11:06:00,395 | INFO | train_grpo_train | metrics_logged mode=train step=98

                                                   
{'loss': 0.0285, 'grad_norm': 9.983766555786133, 'learning_rate': 9.706060606060606e-06, 'num_tokens': 216468.0, 'completions/mean_length': 107.375, 'completions/min_length': 89.0, 'completions/max_length': 135.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 107.375, 'completions/min_terminated_length': 89.0, 'completions/max_terminated_length': 135.0, 'rewards/meter/mean': 0.5628624558448792, 'rewards/meter/std': 0.2919635474681854, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9685242176055908, 'rewards/lexical_plausibility/std': 0.08902685344219208, 'rewards/judge_quality/mean': 0.2562499940395355, 'rewards/judge_quality/std': 0.14744853973388672, 'rewards/repeat_penalty/mean': 0.9908519983291626, 'rewards/repeat_penalty/std': 0.014483519829809666, 'rewards/near_duplicate_penalty/mean': 0.9950829744338989, 'rewards/near_duplicate_penalty/std': 0.004219905938953161, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.995726466178894, 'rewards/distinct_2/std': 0.012087294831871986, 'rewards/total_composite/mean': 0.128617063164711, 'rewards/total_composite/std': 0.08563850075006485, 'reward': 0.128617063164711, 'reward_std': 0.08563850075006485, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20586693286895752, 'sampling/sampling_logp_difference/max': 1.4943318367004395, 'sampling/importance_sampling_ratio/min': 0.22439850866794586, 'sampling/importance_sampling_ratio/mean': 1.0511600971221924, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.7603975236415863, 'clip_ratio/low_mean': 0.10669201333075762, 'clip_ratio/low_min': 0.10669201333075762, 'clip_ratio/high_mean': 0.05855082534253597, 'clip_ratio/high_max': 0.05855082534253597, 'clip_ratio/region_mean': 0.1652428386732936, 'reward_total_mean': 0.128617063164711, 'reward_meter_mean': 0.5628624558448792, 'reward_meter_std': 0.2919635474681854, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9685242176055908, 'reward_lexical_plausibility_std': 0.08902685344219208, 'reward_repeat_penalty_mean': 0.9908519983291626, 'reward_repeat_penalty_std': 0.014483519829809666, 'reward_near_duplicate_penalty_mean': 0.9950829744338989, 'reward_near_duplicate_penalty_std': 0.004219905938953161, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.995726466178894, 'reward_distinct_2_std': 0.012087294831871986, 'reward_judge_quality_mean': 0.2562499940395355, 'reward_judge_quality_std': 0.14744853973388672, 'reward_total_composite_mean': 0.128617063164711, 'reward_total_composite_std': 0.08563850075006485, 'epoch': 0.0}

  3%|▎         | 98/3300 [19:48<9:39:55, 10.87s/it]INFO 04-12 11:06:00 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 99/3300 [19:59<9:38:39, 10.85s/it]2026-04-12 11:06:11,363 | INFO | train_grpo_train | metrics_logged mode=train step=99

                                                   
{'loss': 0.0309, 'grad_norm': 12.82423210144043, 'learning_rate': 9.703030303030305e-06, 'num_tokens': 218558.0, 'completions/mean_length': 93.25, 'completions/min_length': 77.0, 'completions/max_length': 127.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 93.25, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 127.0, 'rewards/meter/mean': 0.8218356370925903, 'rewards/meter/std': 0.23102423548698425, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.9977676272392273, 'rewards/repeat_penalty/std': 0.002441425807774067, 'rewards/near_duplicate_penalty/mean': 0.9977676272392273, 'rewards/near_duplicate_penalty/std': 0.002441425807774067, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.32219815254211426, 'rewards/total_composite/std': 0.10799339413642883, 'reward': 0.32219815254211426, 'reward_std': 0.10799339413642883, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20814354717731476, 'sampling/sampling_logp_difference/max': 1.4589653015136719, 'sampling/importance_sampling_ratio/min': 0.23247671127319336, 'sampling/importance_sampling_ratio/mean': 1.0451229810714722, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.364054650068283, 'clip_ratio/low_mean': 0.09408900327980518, 'clip_ratio/low_min': 0.09408900327980518, 'clip_ratio/high_mean': 0.08094443194568157, 'clip_ratio/high_max': 0.08094443194568157, 'clip_ratio/region_mean': 0.17503343522548676, 'reward_total_mean': 0.32219815254211426, 'reward_meter_mean': 0.8218356370925903, 'reward_meter_std': 0.23102423548698425, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9977676272392273, 'reward_repeat_penalty_std': 0.002441425807774067, 'reward_near_duplicate_penalty_mean': 0.9977676272392273, 'reward_near_duplicate_penalty_std': 0.002441425807774067, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.32219815254211426, 'reward_total_composite_std': 0.10799339413642883, 'epoch': 0.0}

  3%|▎         | 99/3300 [19:59<9:38:39, 10.85s/it]INFO 04-12 11:06:11 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 100/3300 [20:10<9:51:46, 11.10s/it]2026-04-12 11:06:22,871 | INFO | train_grpo_train | metrics_logged mode=train step=100

                                                    
{'loss': 0.0235, 'grad_norm': 14.275720596313477, 'learning_rate': 9.7e-06, 'num_tokens': 220299.0, 'completions/mean_length': 63.625, 'completions/min_length': 56.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.625, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.2841872572898865, 'rewards/meter/std': 0.3190371096134186, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3962499797344208, 'rewards/judge_quality/std': 0.24136148393154144, 'rewards/repeat_penalty/mean': 0.9880180358886719, 'rewards/repeat_penalty/std': 0.010971157811582088, 'rewards/near_duplicate_penalty/mean': 0.9880180358886719, 'rewards/near_duplicate_penalty/std': 0.010971157811582088, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.11206617206335068, 'rewards/total_composite/std': 0.1446806639432907, 'reward': 0.11206617206335068, 'reward_std': 0.1446806639432907, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2242622673511505, 'sampling/sampling_logp_difference/max': 1.5184240341186523, 'sampling/importance_sampling_ratio/min': 0.2190568596124649, 'sampling/importance_sampling_ratio/mean': 1.0265940427780151, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.3064437806606293, 'clip_ratio/low_mean': 0.13668633438646793, 'clip_ratio/low_min': 0.13668633438646793, 'clip_ratio/high_mean': 0.05593487620353699, 'clip_ratio/high_max': 0.05593487620353699, 'clip_ratio/region_mean': 0.19262121059000492, 'reward_total_mean': 0.11206617206335068, 'reward_meter_mean': 0.2841872572898865, 'reward_meter_std': 0.3190371096134186, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9880180358886719, 'reward_repeat_penalty_std': 0.010971157811582088, 'reward_near_duplicate_penalty_mean': 0.9880180358886719, 'reward_near_duplicate_penalty_std': 0.010971157811582088, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3962499797344208, 'reward_judge_quality_std': 0.24136148393154144, 'reward_total_composite_mean': 0.11206617206335068, 'reward_total_composite_std': 0.1446806639432907, 'epoch': 0.0}

  3%|▎         | 100/3300 [20:10<9:51:46, 11.10s/it]INFO 04-12 11:06:23 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


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

 15%|█▌        | 2/13 [00:10<00:55,  5.04s/it]

 23%|██▎       | 3/13 [00:17<00:59,  5.94s/it]

 31%|███       | 4/13 [00:27<01:08,  7.56s/it]

 38%|███▊      | 5/13 [00:39<01:13,  9.21s/it]

 46%|████▌     | 6/13 [00:52<01:12, 10.34s/it]

 54%|█████▍    | 7/13 [01:02<01:01, 10.18s/it]

 62%|██████▏   | 8/13 [01:18<01:00, 12.12s/it]

 69%|██████▉   | 9/13 [01:31<00:48, 12.20s/it]

 77%|███████▋  | 10/13 [01:43<00:36, 12.20s/it]

 85%|████████▍ | 11/13 [01:52<00:22, 11.38s/it]

 92%|█████████▏| 12/13 [02:06<00:12, 12.11s/it]

100%|██████████| 13/13 [02:17<00:00, 11.87s/it]2026-04-12 11:08:52,950 | INFO | train_grpo_train | metrics_logged mode=eval step=100

                                                    

                                               
{'eval_loss': nan, 'eval_runtime': 150.0634, 'eval_samples_per_second': 0.693, 'eval_steps_per_second': 0.087, 'eval_num_tokens': 220299.0, 'eval_completions/mean_length': 179.4903846153846, 'eval_completions/min_length': 41.76923076923077, 'eval_completions/max_length': 391.53846153846155, 'eval_completions/clipped_ratio': 0.019230769230769232, 'eval_completions/mean_terminated_length': 172.60714369553787, 'eval_completions/min_terminated_length': 41.76923076923077, 'eval_completions/max_terminated_length': 360.61538461538464, 'eval_rewards/meter/mean': 0.4438126110113584, 'eval_rewards/meter/std': 0.3760697337297293, 'eval_rewards/count_adherence/mean': 0.9246577987304101, 'eval_rewards/count_adherence/std': 0.08610457095962304, 'eval_rewards/arabic_clean/mean': 0.9038461538461539, 'eval_rewards/arabic_clean/std': 0.21561105434711164, 'eval_rewards/lexical_plausibility/mean': 0.9845591462575473, 'eval_rewards/lexical_plausibility/std': 0.03744004376662465, 'eval_rewards/judge_quality/mean': 0.34538461382572466, 'eval_rewards/judge_quality/std': 0.1808490724517749, 'eval_rewards/repeat_penalty/mean': 0.9725102323752183, 'eval_rewards/repeat_penalty/std': 0.04685022345242592, 'eval_rewards/near_duplicate_penalty/mean': 0.9911677057926471, 'eval_rewards/near_duplicate_penalty/std': 0.009857276257557364, 'eval_rewards/opening_diversity/mean': 0.9927884615384616, 'eval_rewards/opening_diversity/std': 0.020397310646680687, 'eval_rewards/distinct_2/mean': 0.9887164464363685, 'eval_rewards/distinct_2/std': 0.02292853558006195, 'eval_rewards/total_composite/mean': 0.14150516058389956, 'eval_rewards/total_composite/std': 0.1569332918868615, 'eval_reward': 0.14150516058389956, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.15858542231413034, 'eval_sampling/sampling_logp_difference/max': 1.2093834510216346, 'eval_sampling/importance_sampling_ratio/min': 0.3008319517740837, 'eval_sampling/importance_sampling_ratio/mean': 1.0480100833452666, 'eval_sampling/importance_sampling_ratio/max': 1.6585412117151113, 'eval_entropy': 2.6760957057659445, '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': 0.14150516058389956, 'eval_reward_meter_mean': 0.4438126110113584, 'eval_reward_meter_std': 0.3760697337297293, 'eval_reward_count_adherence_mean': 0.9246577987304101, 'eval_reward_count_adherence_std': 0.08610457095962304, 'eval_reward_arabic_clean_mean': 0.9038461538461539, 'eval_reward_arabic_clean_std': 0.21561105434711164, 'eval_reward_lexical_plausibility_mean': 0.9845591462575473, 'eval_reward_lexical_plausibility_std': 0.03744004376662465, 'eval_reward_repeat_penalty_mean': 0.9725102323752183, 'eval_reward_repeat_penalty_std': 0.04685022345242592, 'eval_reward_near_duplicate_penalty_mean': 0.9911677057926471, 'eval_reward_near_duplicate_penalty_std': 0.009857276257557364, 'eval_reward_opening_diversity_mean': 0.9927884615384616, 'eval_reward_opening_diversity_std': 0.020397310646680687, 'eval_reward_distinct_2_mean': 0.9887164464363685, 'eval_reward_distinct_2_std': 0.02292853558006195, 'eval_reward_judge_quality_mean': 0.34538461382572466, 'eval_reward_judge_quality_std': 0.1808490724517749, 'eval_reward_total_composite_mean': 0.14150516058389956, 'eval_reward_total_composite_std': 0.1569332918868615, 'epoch': 0.0}

  3%|▎         | 100/3300 [22:40<9:51:46, 11.10s/it]

100%|██████████| 13/13 [02:17<00:00, 11.87s/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(

  3%|▎         | 101/3300 [22:52<50:05:13, 56.37s/it]2026-04-12 11:09:04,866 | INFO | train_grpo_train | metrics_logged mode=train step=101

                                                     
{'loss': 0.0864, 'grad_norm': 18.904470443725586, 'learning_rate': 9.696969696969698e-06, 'num_tokens': 221926.0, 'completions/mean_length': 48.375, 'completions/min_length': 38.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.375, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.8652718663215637, 'rewards/meter/std': 0.276619553565979, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9791666269302368, 'rewards/lexical_plausibility/std': 0.0589255727827549, 'rewards/judge_quality/mean': 0.4124999940395355, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/repeat_penalty/mean': 0.9877814054489136, 'rewards/repeat_penalty/std': 0.020352782681584358, 'rewards/near_duplicate_penalty/mean': 0.9951778650283813, 'rewards/near_duplicate_penalty/std': 0.007622325327247381, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9926035404205322, 'rewards/distinct_2/std': 0.020920326933264732, 'rewards/total_composite/mean': 0.35213059186935425, 'rewards/total_composite/std': 0.14728707075119019, 'reward': 0.35213059186935425, 'reward_std': 0.14728707075119019, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23319759964942932, 'sampling/sampling_logp_difference/max': 1.485825538635254, 'sampling/importance_sampling_ratio/min': 0.27501600980758667, 'sampling/importance_sampling_ratio/mean': 1.0533989667892456, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.526646763086319, 'clip_ratio/low_mean': 0.027720949612557888, 'clip_ratio/low_min': 0.027720949612557888, 'clip_ratio/high_mean': 0.12845481652766466, 'clip_ratio/high_max': 0.12845481652766466, 'clip_ratio/region_mean': 0.15617576614022255, 'reward_total_mean': 0.35213059186935425, 'reward_meter_mean': 0.8652718663215637, 'reward_meter_std': 0.276619553565979, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9791666269302368, 'reward_lexical_plausibility_std': 0.0589255727827549, 'reward_repeat_penalty_mean': 0.9877814054489136, 'reward_repeat_penalty_std': 0.020352782681584358, 'reward_near_duplicate_penalty_mean': 0.9951778650283813, 'reward_near_duplicate_penalty_std': 0.007622325327247381, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9926035404205322, 'reward_distinct_2_std': 0.020920326933264732, 'reward_judge_quality_mean': 0.4124999940395355, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.35213059186935425, 'reward_total_composite_std': 0.14728707075119019, 'epoch': 0.0}

  3%|▎         | 101/3300 [22:52<50:05:13, 56.37s/it]INFO 04-12 11:09:05 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 102/3300 [23:03<37:59:54, 42.77s/it]2026-04-12 11:09:15,929 | INFO | train_grpo_train | metrics_logged mode=train step=102

                                                     
{'loss': 0.1022, 'grad_norm': 15.38626766204834, 'learning_rate': 9.693939393939395e-06, 'num_tokens': 223500.0, 'completions/mean_length': 45.75, 'completions/min_length': 38.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.75, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.785392701625824, 'rewards/meter/std': 0.3241640627384186, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9870174527168274, 'rewards/lexical_plausibility/std': 0.036720167845487595, 'rewards/judge_quality/mean': 0.6337499618530273, 'rewards/judge_quality/std': 0.24100609123706818, 'rewards/repeat_penalty/mean': 0.9874306917190552, 'rewards/repeat_penalty/std': 0.01784193515777588, 'rewards/near_duplicate_penalty/mean': 0.9874306917190552, 'rewards/near_duplicate_penalty/std': 0.01784193515777588, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.48221316933631897, 'rewards/total_composite/std': 0.3026334345340729, 'reward': 0.48221316933631897, 'reward_std': 0.3026334345340729, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2042502462863922, 'sampling/sampling_logp_difference/max': 1.617997407913208, 'sampling/importance_sampling_ratio/min': 0.19829539954662323, 'sampling/importance_sampling_ratio/mean': 1.0294554233551025, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1104984432458878, 'clip_ratio/low_mean': 0.1281595453619957, 'clip_ratio/low_min': 0.1281595453619957, 'clip_ratio/high_mean': 0.05687452293932438, 'clip_ratio/high_max': 0.05687452293932438, 'clip_ratio/region_mean': 0.18503406830132008, 'reward_total_mean': 0.48221316933631897, 'reward_meter_mean': 0.785392701625824, 'reward_meter_std': 0.3241640627384186, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9870174527168274, 'reward_lexical_plausibility_std': 0.036720167845487595, 'reward_repeat_penalty_mean': 0.9874306917190552, 'reward_repeat_penalty_std': 0.01784193515777588, 'reward_near_duplicate_penalty_mean': 0.9874306917190552, 'reward_near_duplicate_penalty_std': 0.01784193515777588, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6337499618530273, 'reward_judge_quality_std': 0.24100609123706818, 'reward_total_composite_mean': 0.48221316933631897, 'reward_total_composite_std': 0.3026334345340729, 'epoch': 0.0}

  3%|▎         | 102/3300 [23:03<37:59:54, 42.77s/it]INFO 04-12 11:09:16 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 103/3300 [23:13<29:11:47, 32.88s/it]2026-04-12 11:09:25,730 | INFO | train_grpo_train | metrics_logged mode=train step=103

                                                     
{'loss': 0.0183, 'grad_norm': 15.671799659729004, 'learning_rate': 9.690909090909092e-06, 'num_tokens': 224994.0, 'completions/mean_length': 35.75, 'completions/min_length': 32.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.75, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.870998740196228, 'rewards/meter/std': 0.2737838625907898, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9353253841400146, 'rewards/lexical_plausibility/std': 0.1349874883890152, 'rewards/judge_quality/mean': 0.7512500286102295, 'rewards/judge_quality/std': 0.24976776540279388, 'rewards/repeat_penalty/mean': 0.9875068068504333, 'rewards/repeat_penalty/std': 0.018238745629787445, 'rewards/near_duplicate_penalty/mean': 0.9875068068504333, 'rewards/near_duplicate_penalty/std': 0.018238745629787445, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6357969045639038, 'rewards/total_composite/std': 0.29190483689308167, 'reward': 0.6357969045639038, 'reward_std': 0.2919048070907593, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1651766300201416, 'sampling/sampling_logp_difference/max': 2.670238494873047, 'sampling/importance_sampling_ratio/min': 0.06923571228981018, 'sampling/importance_sampling_ratio/mean': 1.0023692846298218, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3214549124240875, 'clip_ratio/low_mean': 0.06165114138275385, 'clip_ratio/low_min': 0.06165114138275385, 'clip_ratio/high_mean': 0.09113431721925735, 'clip_ratio/high_max': 0.09113431721925735, 'clip_ratio/region_mean': 0.1527854586020112, 'reward_total_mean': 0.6357969045639038, 'reward_meter_mean': 0.870998740196228, 'reward_meter_std': 0.2737838625907898, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9353253841400146, 'reward_lexical_plausibility_std': 0.1349874883890152, 'reward_repeat_penalty_mean': 0.9875068068504333, 'reward_repeat_penalty_std': 0.018238745629787445, 'reward_near_duplicate_penalty_mean': 0.9875068068504333, 'reward_near_duplicate_penalty_std': 0.018238745629787445, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7512500286102295, 'reward_judge_quality_std': 0.24976776540279388, 'reward_total_composite_mean': 0.6357969045639038, 'reward_total_composite_std': 0.29190483689308167, 'epoch': 0.0}

  3%|▎         | 103/3300 [23:13<29:11:47, 32.88s/it]INFO 04-12 11:09:25 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 104/3300 [23:23<23:01:47, 25.94s/it]2026-04-12 11:09:35,473 | INFO | train_grpo_train | metrics_logged mode=train step=104

                                                     
{'loss': 0.0704, 'grad_norm': 20.590539932250977, 'learning_rate': 9.687878787878788e-06, 'num_tokens': 226400.0, 'completions/mean_length': 19.75, 'completions/min_length': 15.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.75, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.8273722529411316, 'rewards/meter/std': 0.33302101492881775, '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/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.5087499618530273, 'rewards/judge_quality/std': 0.37288209795951843, 'rewards/repeat_penalty/mean': 0.7364565134048462, 'rewards/repeat_penalty/std': 0.025085056200623512, 'rewards/near_duplicate_penalty/mean': 0.9819420576095581, 'rewards/near_duplicate_penalty/std': 0.033446744084358215, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.44802480936050415, 'rewards/total_composite/std': 0.4030570089817047, 'reward': 0.44802480936050415, 'reward_std': 0.4030570089817047, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2312489151954651, 'sampling/sampling_logp_difference/max': 1.2237434387207031, 'sampling/importance_sampling_ratio/min': 0.29412707686424255, 'sampling/importance_sampling_ratio/mean': 1.0077565908432007, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1665150076150894, 'clip_ratio/low_mean': 0.11906194407492876, 'clip_ratio/low_min': 0.11906194407492876, 'clip_ratio/high_mean': 0.06388888973742723, 'clip_ratio/high_max': 0.06388888973742723, 'clip_ratio/region_mean': 0.182950833812356, 'reward_total_mean': 0.44802480936050415, 'reward_meter_mean': 0.8273722529411316, 'reward_meter_std': 0.33302101492881775, '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_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.7364565134048462, 'reward_repeat_penalty_std': 0.025085056200623512, 'reward_near_duplicate_penalty_mean': 0.9819420576095581, 'reward_near_duplicate_penalty_std': 0.033446744084358215, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5087499618530273, 'reward_judge_quality_std': 0.37288209795951843, 'reward_total_composite_mean': 0.44802480936050415, 'reward_total_composite_std': 0.4030570089817047, 'epoch': 0.0}

  3%|▎         | 104/3300 [23:23<23:01:47, 25.94s/it]INFO 04-12 11:09:35 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 105/3300 [23:32<18:35:28, 20.95s/it]2026-04-12 11:09:44,768 | INFO | train_grpo_train | metrics_logged mode=train step=105

                                                     
{'loss': 0.014, 'grad_norm': 12.693693161010742, 'learning_rate': 9.684848484848487e-06, 'num_tokens': 228284.0, 'completions/mean_length': 69.5, 'completions/min_length': 41.0, 'completions/max_length': 95.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.5, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 95.0, 'rewards/meter/mean': 0.5121266841888428, 'rewards/meter/std': 0.34023505449295044, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9752435088157654, 'rewards/lexical_plausibility/std': 0.046044979244470596, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.05345224589109421, 'rewards/repeat_penalty/mean': 0.9891045689582825, 'rewards/repeat_penalty/std': 0.013739429414272308, 'rewards/near_duplicate_penalty/mean': 0.9891045689582825, 'rewards/near_duplicate_penalty/std': 0.013739429414272308, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2042568027973175, 'rewards/total_composite/std': 0.139831081032753, 'reward': 0.2042568027973175, 'reward_std': 0.1398310661315918, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2049107700586319, 'sampling/sampling_logp_difference/max': 1.7144489288330078, 'sampling/importance_sampling_ratio/min': 0.1800629198551178, 'sampling/importance_sampling_ratio/mean': 1.034571886062622, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0561163872480392, 'clip_ratio/low_mean': 0.0917549766600132, 'clip_ratio/low_min': 0.0917549766600132, 'clip_ratio/high_mean': 0.11717102490365505, 'clip_ratio/high_max': 0.11717102490365505, 'clip_ratio/region_mean': 0.20892600156366825, 'reward_total_mean': 0.2042568027973175, 'reward_meter_mean': 0.5121266841888428, 'reward_meter_std': 0.34023505449295044, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9752435088157654, 'reward_lexical_plausibility_std': 0.046044979244470596, 'reward_repeat_penalty_mean': 0.9891045689582825, 'reward_repeat_penalty_std': 0.013739429414272308, 'reward_near_duplicate_penalty_mean': 0.9891045689582825, 'reward_near_duplicate_penalty_std': 0.013739429414272308, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.05345224589109421, 'reward_total_composite_mean': 0.2042568027973175, 'reward_total_composite_std': 0.139831081032753, 'epoch': 0.0}

  3%|▎         | 105/3300 [23:32<18:35:28, 20.95s/it]INFO 04-12 11:09:45 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 106/3300 [23:41<15:16:01, 17.21s/it]2026-04-12 11:09:53,246 | INFO | train_grpo_train | metrics_logged mode=train step=106

                                                     
{'loss': 0.0802, 'grad_norm': 19.78082275390625, 'learning_rate': 9.681818181818182e-06, 'num_tokens': 230063.0, 'completions/mean_length': 37.375, 'completions/min_length': 32.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.375, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.4091936945915222, 'rewards/meter/std': 0.38513103127479553, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3712500035762787, 'rewards/judge_quality/std': 0.24532414972782135, 'rewards/repeat_penalty/mean': 0.9900262355804443, 'rewards/repeat_penalty/std': 0.019093554466962814, 'rewards/near_duplicate_penalty/mean': 0.9900262355804443, 'rewards/near_duplicate_penalty/std': 0.019093554466962814, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.12962524592876434, 'rewards/total_composite/std': 0.12251812219619751, 'reward': 0.12962524592876434, 'reward_std': 0.12251812219619751, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19340533018112183, 'sampling/sampling_logp_difference/max': 1.4765138626098633, 'sampling/importance_sampling_ratio/min': 0.22843265533447266, 'sampling/importance_sampling_ratio/mean': 1.0042545795440674, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3978212103247643, 'clip_ratio/low_mean': 0.10965030081570148, 'clip_ratio/low_min': 0.10965030081570148, 'clip_ratio/high_mean': 0.08934779465198517, 'clip_ratio/high_max': 0.08934779465198517, 'clip_ratio/region_mean': 0.19899809546768665, 'reward_total_mean': 0.12962524592876434, 'reward_meter_mean': 0.4091936945915222, 'reward_meter_std': 0.38513103127479553, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9900262355804443, 'reward_repeat_penalty_std': 0.019093554466962814, 'reward_near_duplicate_penalty_mean': 0.9900262355804443, 'reward_near_duplicate_penalty_std': 0.019093554466962814, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3712500035762787, 'reward_judge_quality_std': 0.24532414972782135, 'reward_total_composite_mean': 0.12962524592876434, 'reward_total_composite_std': 0.12251812219619751, 'epoch': 0.0}

  3%|▎         | 106/3300 [23:41<15:16:01, 17.21s/it]INFO 04-12 11:09:53 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 107/3300 [23:51<13:29:48, 15.22s/it]2026-04-12 11:10:03,958 | INFO | train_grpo_train | metrics_logged mode=train step=107

                                                     
{'loss': 0.1234, 'grad_norm': 12.042773246765137, 'learning_rate': 9.67878787878788e-06, 'num_tokens': 231925.0, 'completions/mean_length': 67.75, 'completions/min_length': 54.0, 'completions/max_length': 100.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.75, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 100.0, 'rewards/meter/mean': 0.8805147409439087, 'rewards/meter/std': 0.19566473364830017, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9791666865348816, 'rewards/lexical_plausibility/std': 0.038575831800699234, 'rewards/judge_quality/mean': 0.36249998211860657, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.9657440185546875, 'rewards/repeat_penalty/std': 0.044199053198099136, 'rewards/near_duplicate_penalty/mean': 0.977892279624939, 'rewards/near_duplicate_penalty/std': 0.024622080847620964, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9871795177459717, 'rewards/distinct_2/std': 0.023738976567983627, 'rewards/total_composite/mean': 0.32012301683425903, 'rewards/total_composite/std': 0.1213817298412323, 'reward': 0.32012301683425903, 'reward_std': 0.1213817223906517, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23260825872421265, 'sampling/sampling_logp_difference/max': 1.4933059215545654, 'sampling/importance_sampling_ratio/min': 0.22462882101535797, 'sampling/importance_sampling_ratio/mean': 1.0501987934112549, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6822806298732758, 'clip_ratio/low_mean': 0.08087963052093983, 'clip_ratio/low_min': 0.08087963052093983, 'clip_ratio/high_mean': 0.14885154739022255, 'clip_ratio/high_max': 0.14885154739022255, 'clip_ratio/region_mean': 0.22973117791116238, 'reward_total_mean': 0.32012301683425903, 'reward_meter_mean': 0.8805147409439087, 'reward_meter_std': 0.19566473364830017, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9791666865348816, 'reward_lexical_plausibility_std': 0.038575831800699234, 'reward_repeat_penalty_mean': 0.9657440185546875, 'reward_repeat_penalty_std': 0.044199053198099136, 'reward_near_duplicate_penalty_mean': 0.977892279624939, 'reward_near_duplicate_penalty_std': 0.024622080847620964, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9871795177459717, 'reward_distinct_2_std': 0.023738976567983627, 'reward_judge_quality_mean': 0.36249998211860657, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.32012301683425903, 'reward_total_composite_std': 0.1213817298412323, 'epoch': 0.0}

  3%|▎         | 107/3300 [23:51<13:29:48, 15.22s/it]INFO 04-12 11:10:04 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 108/3300 [24:04<12:46:47, 14.41s/it]2026-04-12 11:10:16,364 | INFO | train_grpo_train | metrics_logged mode=train step=108

                                                     
{'loss': -0.0024, 'grad_norm': 6.266803741455078, 'learning_rate': 9.675757575757577e-06, 'num_tokens': 235644.0, 'completions/mean_length': 264.875, 'completions/min_length': 185.0, 'completions/max_length': 360.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 264.875, 'completions/min_terminated_length': 185.0, 'completions/max_terminated_length': 360.0, 'rewards/meter/mean': 0.195610910654068, 'rewards/meter/std': 0.12062400579452515, 'rewards/count_adherence/mean': 0.8375000357627869, 'rewards/count_adherence/std': 0.07440237700939178, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.9957386255264282, 'rewards/lexical_plausibility/std': 0.012052967213094234, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.10690449178218842, 'rewards/repeat_penalty/mean': 0.9829477071762085, 'rewards/repeat_penalty/std': 0.01583346351981163, 'rewards/near_duplicate_penalty/mean': 0.991424024105072, 'rewards/near_duplicate_penalty/std': 0.006038422230631113, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9914255142211914, 'rewards/distinct_2/std': 0.012757115066051483, 'rewards/total_composite/mean': 0.030561331659555435, 'rewards/total_composite/std': 0.04264984652400017, 'reward': 0.030561331659555435, 'reward_std': 0.04264984279870987, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23715300858020782, 'sampling/sampling_logp_difference/max': 1.967503547668457, 'sampling/importance_sampling_ratio/min': 0.13980543613433838, 'sampling/importance_sampling_ratio/mean': 1.0414592027664185, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.3490786850452423, 'clip_ratio/low_mean': 0.1096208244562149, 'clip_ratio/low_min': 0.1096208244562149, 'clip_ratio/high_mean': 0.08376534841954708, 'clip_ratio/high_max': 0.08376534841954708, 'clip_ratio/region_mean': 0.19338617287576199, 'reward_total_mean': 0.030561331659555435, 'reward_meter_mean': 0.195610910654068, 'reward_meter_std': 0.12062400579452515, 'reward_count_adherence_mean': 0.8375000357627869, 'reward_count_adherence_std': 0.07440237700939178, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.9957386255264282, 'reward_lexical_plausibility_std': 0.012052967213094234, 'reward_repeat_penalty_mean': 0.9829477071762085, 'reward_repeat_penalty_std': 0.01583346351981163, 'reward_near_duplicate_penalty_mean': 0.991424024105072, 'reward_near_duplicate_penalty_std': 0.006038422230631113, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9914255142211914, 'reward_distinct_2_std': 0.012757115066051483, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.10690449178218842, 'reward_total_composite_mean': 0.030561331659555435, 'reward_total_composite_std': 0.04264984652400017, 'epoch': 0.0}

  3%|▎         | 108/3300 [24:04<12:46:47, 14.41s/it]INFO 04-12 11:10:16 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 109/3300 [24:12<11:05:25, 12.51s/it]2026-04-12 11:10:24,443 | INFO | train_grpo_train | metrics_logged mode=train step=109

                                                     
{'loss': 0.181, 'grad_norm': 19.989253997802734, 'learning_rate': 9.672727272727274e-06, 'num_tokens': 237016.0, 'completions/mean_length': 24.5, 'completions/min_length': 17.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.5, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.630812406539917, 'rewards/meter/std': 0.452547162771225, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8837499618530273, 'rewards/judge_quality/std': 0.17557558417320251, 'rewards/repeat_penalty/mean': 0.7260398864746094, 'rewards/repeat_penalty/std': 0.04937082529067993, 'rewards/near_duplicate_penalty/mean': 0.9680531620979309, 'rewards/near_duplicate_penalty/std': 0.06582777202129364, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.537653923034668, 'rewards/total_composite/std': 0.40953513979911804, 'reward': 0.537653923034668, 'reward_std': 0.40953510999679565, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14390215277671814, 'sampling/sampling_logp_difference/max': 1.2209405899047852, 'sampling/importance_sampling_ratio/min': 0.2949526011943817, 'sampling/importance_sampling_ratio/mean': 1.0361294746398926, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3929248750209808, 'clip_ratio/low_mean': 0.08533278107643127, 'clip_ratio/low_min': 0.08533278107643127, 'clip_ratio/high_mean': 0.07347927987575531, 'clip_ratio/high_max': 0.07347927987575531, 'clip_ratio/region_mean': 0.15881206095218658, 'reward_total_mean': 0.537653923034668, 'reward_meter_mean': 0.630812406539917, 'reward_meter_std': 0.452547162771225, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7260398864746094, 'reward_repeat_penalty_std': 0.04937082529067993, 'reward_near_duplicate_penalty_mean': 0.9680531620979309, 'reward_near_duplicate_penalty_std': 0.06582777202129364, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8837499618530273, 'reward_judge_quality_std': 0.17557558417320251, 'reward_total_composite_mean': 0.537653923034668, 'reward_total_composite_std': 0.40953513979911804, 'epoch': 0.0}

  3%|▎         | 109/3300 [24:12<11:05:25, 12.51s/it]INFO 04-12 11:10:24 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 110/3300 [24:21<10:06:39, 11.41s/it]2026-04-12 11:10:33,273 | INFO | train_grpo_train | metrics_logged mode=train step=110

                                                     
{'loss': 0.1281, 'grad_norm': 13.449333190917969, 'learning_rate': 9.66969696969697e-06, 'num_tokens': 238652.0, 'completions/mean_length': 49.5, 'completions/min_length': 34.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.5, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.837989866733551, 'rewards/meter/std': 0.23016050457954407, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.4874999523162842, 'rewards/judge_quality/std': 0.14078859984874725, 'rewards/repeat_penalty/mean': 0.9668082594871521, 'rewards/repeat_penalty/std': 0.030257586389780045, 'rewards/near_duplicate_penalty/mean': 0.9668082594871521, 'rewards/near_duplicate_penalty/std': 0.030257586389780045, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3983583450317383, 'rewards/total_composite/std': 0.15410736203193665, 'reward': 0.3983583450317383, 'reward_std': 0.15410736203193665, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18442291021347046, 'sampling/sampling_logp_difference/max': 1.480802059173584, 'sampling/importance_sampling_ratio/min': 0.22745519876480103, 'sampling/importance_sampling_ratio/mean': 1.03956139087677, 'sampling/importance_sampling_ratio/max': 1.9880635738372803, 'entropy': 2.166471987962723, 'clip_ratio/low_mean': 0.09994436614215374, 'clip_ratio/low_min': 0.09994436614215374, 'clip_ratio/high_mean': 0.036318263970315456, 'clip_ratio/high_max': 0.036318263970315456, 'clip_ratio/region_mean': 0.1362626301124692, 'reward_total_mean': 0.3983583450317383, 'reward_meter_mean': 0.837989866733551, 'reward_meter_std': 0.23016050457954407, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.9668082594871521, 'reward_repeat_penalty_std': 0.030257586389780045, 'reward_near_duplicate_penalty_mean': 0.9668082594871521, 'reward_near_duplicate_penalty_std': 0.030257586389780045, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4874999523162842, 'reward_judge_quality_std': 0.14078859984874725, 'reward_total_composite_mean': 0.3983583450317383, 'reward_total_composite_std': 0.15410736203193665, 'epoch': 0.0}

  3%|▎         | 110/3300 [24:21<10:06:39, 11.41s/it]INFO 04-12 11:10:33 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 111/3300 [24:29<9:22:29, 10.58s/it] 2026-04-12 11:10:41,926 | INFO | train_grpo_train | metrics_logged mode=train step=111

                                                    
{'loss': 0.1662, 'grad_norm': 21.777551651000977, 'learning_rate': 9.666666666666667e-06, 'num_tokens': 240079.0, 'completions/mean_length': 23.375, 'completions/min_length': 17.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.375, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.6233583092689514, 'rewards/meter/std': 0.4267948269844055, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9137561321258545, 'rewards/lexical_plausibility/std': 0.1420510709285736, 'rewards/judge_quality/mean': 0.6762499809265137, 'rewards/judge_quality/std': 0.3651198446750641, 'rewards/repeat_penalty/mean': 0.740288496017456, 'rewards/repeat_penalty/std': 0.019515199586749077, 'rewards/near_duplicate_penalty/mean': 0.9870513677597046, 'rewards/near_duplicate_penalty/std': 0.026020271703600883, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.450778067111969, 'rewards/total_composite/std': 0.40946274995803833, 'reward': 0.450778067111969, 'reward_std': 0.40946274995803833, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1604246199131012, 'sampling/sampling_logp_difference/max': 1.4545950889587402, 'sampling/importance_sampling_ratio/min': 0.23349489271640778, 'sampling/importance_sampling_ratio/mean': 1.012116551399231, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1280418634414673, 'clip_ratio/low_mean': 0.04994176980108023, 'clip_ratio/low_min': 0.04994176980108023, 'clip_ratio/high_mean': 0.05896358657628298, 'clip_ratio/high_max': 0.05896358657628298, 'clip_ratio/region_mean': 0.1089053563773632, 'reward_total_mean': 0.450778067111969, 'reward_meter_mean': 0.6233583092689514, 'reward_meter_std': 0.4267948269844055, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9137561321258545, 'reward_lexical_plausibility_std': 0.1420510709285736, 'reward_repeat_penalty_mean': 0.740288496017456, 'reward_repeat_penalty_std': 0.019515199586749077, 'reward_near_duplicate_penalty_mean': 0.9870513677597046, 'reward_near_duplicate_penalty_std': 0.026020271703600883, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6762499809265137, 'reward_judge_quality_std': 0.3651198446750641, 'reward_total_composite_mean': 0.450778067111969, 'reward_total_composite_std': 0.40946274995803833, 'epoch': 0.0}

  3%|▎         | 111/3300 [24:29<9:22:29, 10.58s/it]INFO 04-12 11:10:42 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 112/3300 [24:38<8:53:40, 10.04s/it]2026-04-12 11:10:50,715 | INFO | train_grpo_train | metrics_logged mode=train step=112

                                                    
{'loss': 0.0108, 'grad_norm': 16.79600715637207, 'learning_rate': 9.663636363636364e-06, 'num_tokens': 241638.0, 'completions/mean_length': 41.875, 'completions/min_length': 35.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.875, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.7554359436035156, 'rewards/meter/std': 0.31206610798835754, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9822304248809814, 'rewards/lexical_plausibility/std': 0.0502600334584713, 'rewards/judge_quality/mean': 0.4712499976158142, 'rewards/judge_quality/std': 0.20951901376247406, 'rewards/repeat_penalty/mean': 0.9948287010192871, 'rewards/repeat_penalty/std': 0.007565564475953579, 'rewards/near_duplicate_penalty/mean': 0.9948287010192871, 'rewards/near_duplicate_penalty/std': 0.007565564475953579, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3201366364955902, 'rewards/total_composite/std': 0.15204645693302155, 'reward': 0.3201366364955902, 'reward_std': 0.15204645693302155, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20717382431030273, 'sampling/sampling_logp_difference/max': 1.4654204845428467, 'sampling/importance_sampling_ratio/min': 0.23098085820674896, 'sampling/importance_sampling_ratio/mean': 1.0283093452453613, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.477309077978134, 'clip_ratio/low_mean': 0.08162708580493927, 'clip_ratio/low_min': 0.08162708580493927, 'clip_ratio/high_mean': 0.15361600555479527, 'clip_ratio/high_max': 0.15361600555479527, 'clip_ratio/region_mean': 0.23524309135973454, 'reward_total_mean': 0.3201366364955902, 'reward_meter_mean': 0.7554359436035156, 'reward_meter_std': 0.31206610798835754, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9822304248809814, 'reward_lexical_plausibility_std': 0.0502600334584713, 'reward_repeat_penalty_mean': 0.9948287010192871, 'reward_repeat_penalty_std': 0.007565564475953579, 'reward_near_duplicate_penalty_mean': 0.9948287010192871, 'reward_near_duplicate_penalty_std': 0.007565564475953579, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4712499976158142, 'reward_judge_quality_std': 0.20951901376247406, 'reward_total_composite_mean': 0.3201366364955902, 'reward_total_composite_std': 0.15204645693302155, 'epoch': 0.0}

  3%|▎         | 112/3300 [24:38<8:53:40, 10.04s/it]INFO 04-12 11:10:50 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 113/3300 [24:48<8:54:13, 10.06s/it]2026-04-12 11:11:00,827 | INFO | train_grpo_train | metrics_logged mode=train step=113

                                                    
{'loss': 0.0827, 'grad_norm': 12.689526557922363, 'learning_rate': 9.660606060606061e-06, 'num_tokens': 243549.0, 'completions/mean_length': 72.875, 'completions/min_length': 60.0, 'completions/max_length': 86.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.875, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.6673098802566528, 'rewards/meter/std': 0.34351009130477905, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9902341961860657, 'rewards/lexical_plausibility/std': 0.02134176902472973, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.9952738285064697, 'rewards/repeat_penalty/std': 0.005715993698686361, 'rewards/near_duplicate_penalty/mean': 0.9952738285064697, 'rewards/near_duplicate_penalty/std': 0.005715993698686361, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.25624915957450867, 'rewards/total_composite/std': 0.16143901646137238, 'reward': 0.25624915957450867, 'reward_std': 0.16143900156021118, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22147944569587708, 'sampling/sampling_logp_difference/max': 2.659895896911621, 'sampling/importance_sampling_ratio/min': 0.0699555054306984, 'sampling/importance_sampling_ratio/mean': 1.0350232124328613, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0763853788375854, 'clip_ratio/low_mean': 0.06416316702961922, 'clip_ratio/low_min': 0.06416316702961922, 'clip_ratio/high_mean': 0.1158969309180975, 'clip_ratio/high_max': 0.1158969309180975, 'clip_ratio/region_mean': 0.1800600979477167, 'reward_total_mean': 0.25624915957450867, 'reward_meter_mean': 0.6673098802566528, 'reward_meter_std': 0.34351009130477905, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9902341961860657, 'reward_lexical_plausibility_std': 0.02134176902472973, 'reward_repeat_penalty_mean': 0.9952738285064697, 'reward_repeat_penalty_std': 0.005715993698686361, 'reward_near_duplicate_penalty_mean': 0.9952738285064697, 'reward_near_duplicate_penalty_std': 0.005715993698686361, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.25624915957450867, 'reward_total_composite_std': 0.16143901646137238, 'epoch': 0.0}

  3%|▎         | 113/3300 [24:48<8:54:13, 10.06s/it]INFO 04-12 11:11:01 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 114/3300 [24:58<8:46:41,  9.92s/it]2026-04-12 11:11:10,565 | INFO | train_grpo_train | metrics_logged mode=train step=114

                                                    
{'loss': 0.094, 'grad_norm': 13.747782707214355, 'learning_rate': 9.657575757575758e-06, 'num_tokens': 245461.0, 'completions/mean_length': 70.0, 'completions/min_length': 51.0, 'completions/max_length': 97.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.0, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 97.0, 'rewards/meter/mean': 0.5687057971954346, 'rewards/meter/std': 0.3622046411037445, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4624999761581421, 'rewards/judge_quality/std': 0.20310096442699432, 'rewards/repeat_penalty/mean': 0.9959641098976135, 'rewards/repeat_penalty/std': 0.0034096972085535526, 'rewards/near_duplicate_penalty/mean': 0.9959641098976135, 'rewards/near_duplicate_penalty/std': 0.0034096972085535526, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2490294873714447, 'rewards/total_composite/std': 0.1855943202972412, 'reward': 0.2490294873714447, 'reward_std': 0.18559430539608002, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20971868932247162, 'sampling/sampling_logp_difference/max': 1.607193946838379, 'sampling/importance_sampling_ratio/min': 0.20044930279254913, 'sampling/importance_sampling_ratio/mean': 1.0306721925735474, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.081496089696884, 'clip_ratio/low_mean': 0.06832332257181406, 'clip_ratio/low_min': 0.06832332257181406, 'clip_ratio/high_mean': 0.11153036169707775, 'clip_ratio/high_max': 0.11153036169707775, 'clip_ratio/region_mean': 0.1798536842688918, 'reward_total_mean': 0.2490294873714447, 'reward_meter_mean': 0.5687057971954346, 'reward_meter_std': 0.3622046411037445, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9959641098976135, 'reward_repeat_penalty_std': 0.0034096972085535526, 'reward_near_duplicate_penalty_mean': 0.9959641098976135, 'reward_near_duplicate_penalty_std': 0.0034096972085535526, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4624999761581421, 'reward_judge_quality_std': 0.20310096442699432, 'reward_total_composite_mean': 0.2490294873714447, 'reward_total_composite_std': 0.1855943202972412, 'epoch': 0.0}

  3%|▎         | 114/3300 [24:58<8:46:41,  9.92s/it]INFO 04-12 11:11:10 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 115/3300 [25:07<8:31:25,  9.63s/it]2026-04-12 11:11:19,443 | INFO | train_grpo_train | metrics_logged mode=train step=115

                                                    
{'loss': 0.0328, 'grad_norm': 19.110523223876953, 'learning_rate': 9.654545454545456e-06, 'num_tokens': 246852.0, 'completions/mean_length': 33.875, 'completions/min_length': 29.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.875, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.4572555422782898, 'rewards/meter/std': 0.4458775818347931, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9634564518928528, 'rewards/lexical_plausibility/std': 0.10336080938577652, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.05345224589109421, 'rewards/repeat_penalty/mean': 0.9571646451950073, 'rewards/repeat_penalty/std': 0.05079825595021248, 'rewards/near_duplicate_penalty/mean': 0.9774123430252075, 'rewards/near_duplicate_penalty/std': 0.03181653469800949, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9791947603225708, 'rewards/distinct_2/std': 0.03854253515601158, 'rewards/total_composite/mean': 0.17574816942214966, 'rewards/total_composite/std': 0.17165198922157288, 'reward': 0.17574816942214966, 'reward_std': 0.17165197432041168, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2393278032541275, 'sampling/sampling_logp_difference/max': 1.797713279724121, 'sampling/importance_sampling_ratio/min': 0.16567730903625488, 'sampling/importance_sampling_ratio/mean': 1.054547905921936, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.7014783918857574, 'clip_ratio/low_mean': 0.10610019788146019, 'clip_ratio/low_min': 0.10610019788146019, 'clip_ratio/high_mean': 0.06926235742866993, 'clip_ratio/high_max': 0.06926235742866993, 'clip_ratio/region_mean': 0.17536255531013012, 'reward_total_mean': 0.17574816942214966, 'reward_meter_mean': 0.4572555422782898, 'reward_meter_std': 0.4458775818347931, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9634564518928528, 'reward_lexical_plausibility_std': 0.10336080938577652, 'reward_repeat_penalty_mean': 0.9571646451950073, 'reward_repeat_penalty_std': 0.05079825595021248, 'reward_near_duplicate_penalty_mean': 0.9774123430252075, 'reward_near_duplicate_penalty_std': 0.03181653469800949, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9791947603225708, 'reward_distinct_2_std': 0.03854253515601158, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.05345224589109421, 'reward_total_composite_mean': 0.17574816942214966, 'reward_total_composite_std': 0.17165198922157288, 'epoch': 0.0}

  3%|▎         | 115/3300 [25:07<8:31:25,  9.63s/it]INFO 04-12 11:11:19 [block_pool.py:378] Successfully reset prefix cache

  4%|▎         | 116/3300 [25:19<9:17:51, 10.51s/it]2026-04-12 11:11:31,930 | INFO | train_grpo_train | metrics_logged mode=train step=116

                                                    
{'loss': 0.1184, 'grad_norm': 7.6971516609191895, 'learning_rate': 9.651515151515153e-06, 'num_tokens': 249893.0, 'completions/mean_length': 158.125, 'completions/min_length': 107.0, 'completions/max_length': 238.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 158.125, 'completions/min_terminated_length': 107.0, 'completions/max_terminated_length': 238.0, 'rewards/meter/mean': 0.32173824310302734, 'rewards/meter/std': 0.3033269941806793, 'rewards/count_adherence/mean': 0.8958333134651184, 'rewards/count_adherence/std': 0.12400396913290024, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9779738187789917, 'rewards/lexical_plausibility/std': 0.03407338261604309, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.984792947769165, 'rewards/repeat_penalty/std': 0.011711930856108665, 'rewards/near_duplicate_penalty/mean': 0.9903175830841064, 'rewards/near_duplicate_penalty/std': 0.004368250258266926, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.994415283203125, 'rewards/distinct_2/std': 0.010353867895901203, 'rewards/total_composite/mean': 0.10354895889759064, 'rewards/total_composite/std': 0.1389368623495102, 'reward': 0.10354895889759064, 'reward_std': 0.1389368623495102, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22458046674728394, 'sampling/sampling_logp_difference/max': 2.5895376205444336, 'sampling/importance_sampling_ratio/min': 0.07505473494529724, 'sampling/importance_sampling_ratio/mean': 1.0412689447402954, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.583432048559189, 'clip_ratio/low_mean': 0.1137649267911911, 'clip_ratio/low_min': 0.1137649267911911, 'clip_ratio/high_mean': 0.06895342469215393, 'clip_ratio/high_max': 0.06895342469215393, 'clip_ratio/region_mean': 0.18271835148334503, 'reward_total_mean': 0.10354895889759064, 'reward_meter_mean': 0.32173824310302734, 'reward_meter_std': 0.3033269941806793, 'reward_count_adherence_mean': 0.8958333134651184, 'reward_count_adherence_std': 0.12400396913290024, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9779738187789917, 'reward_lexical_plausibility_std': 0.03407338261604309, 'reward_repeat_penalty_mean': 0.984792947769165, 'reward_repeat_penalty_std': 0.011711930856108665, 'reward_near_duplicate_penalty_mean': 0.9903175830841064, 'reward_near_duplicate_penalty_std': 0.004368250258266926, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.994415283203125, 'reward_distinct_2_std': 0.010353867895901203, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.10354895889759064, 'reward_total_composite_std': 0.1389368623495102, 'epoch': 0.0}

  4%|▎         | 116/3300 [25:19<9:17:51, 10.51s/it]INFO 04-12 11:11:32 [block_pool.py:378] Successfully reset prefix cache

  4%|▎         | 117/3300 [25:28<8:53:52, 10.06s/it]2026-04-12 11:11:40,950 | INFO | train_grpo_train | metrics_logged mode=train step=117

                                                    
{'loss': 0.0382, 'grad_norm': 13.78189468383789, 'learning_rate': 9.648484848484849e-06, 'num_tokens': 251506.0, 'completions/mean_length': 42.625, 'completions/min_length': 31.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.625, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.49554315209388733, 'rewards/meter/std': 0.44302481412887573, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9687139391899109, 'rewards/lexical_plausibility/std': 0.08849029988050461, 'rewards/judge_quality/mean': 0.5137500166893005, 'rewards/judge_quality/std': 0.3464488387107849, 'rewards/repeat_penalty/mean': 0.9968554973602295, 'rewards/repeat_penalty/std': 0.006410405971109867, 'rewards/near_duplicate_penalty/mean': 0.9968554973602295, 'rewards/near_duplicate_penalty/std': 0.006410405971109867, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2768799364566803, 'rewards/total_composite/std': 0.3126735985279083, 'reward': 0.2768799364566803, 'reward_std': 0.3126735985279083, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19522778689861298, 'sampling/sampling_logp_difference/max': 1.3107595443725586, 'sampling/importance_sampling_ratio/min': 0.26961517333984375, 'sampling/importance_sampling_ratio/mean': 1.0474461317062378, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.915073737502098, 'clip_ratio/low_mean': 0.11497455555945635, 'clip_ratio/low_min': 0.11497455555945635, 'clip_ratio/high_mean': 0.08457787334918976, 'clip_ratio/high_max': 0.08457787334918976, 'clip_ratio/region_mean': 0.1995524289086461, 'reward_total_mean': 0.2768799364566803, 'reward_meter_mean': 0.49554315209388733, 'reward_meter_std': 0.44302481412887573, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9687139391899109, 'reward_lexical_plausibility_std': 0.08849029988050461, 'reward_repeat_penalty_mean': 0.9968554973602295, 'reward_repeat_penalty_std': 0.006410405971109867, 'reward_near_duplicate_penalty_mean': 0.9968554973602295, 'reward_near_duplicate_penalty_std': 0.006410405971109867, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5137500166893005, 'reward_judge_quality_std': 0.3464488387107849, 'reward_total_composite_mean': 0.2768799364566803, 'reward_total_composite_std': 0.3126735985279083, 'epoch': 0.0}

  4%|▎         | 117/3300 [25:28<8:53:52, 10.06s/it]INFO 04-12 11:11:41 [block_pool.py:378] Successfully reset prefix cache

  4%|▎         | 118/3300 [25:38<8:46:46,  9.93s/it]2026-04-12 11:11:50,574 | INFO | train_grpo_train | metrics_logged mode=train step=118

                                                    
{'loss': 0.0299, 'grad_norm': 11.150492668151855, 'learning_rate': 9.645454545454548e-06, 'num_tokens': 253279.0, 'completions/mean_length': 48.625, 'completions/min_length': 38.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.625, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.7285923957824707, 'rewards/meter/std': 0.3594067692756653, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9673515558242798, 'rewards/lexical_plausibility/std': 0.09234381467103958, 'rewards/judge_quality/mean': 0.6262500286102295, 'rewards/judge_quality/std': 0.2432481348514557, 'rewards/repeat_penalty/mean': 0.9713559150695801, 'rewards/repeat_penalty/std': 0.07632871717214584, 'rewards/near_duplicate_penalty/mean': 0.9891414642333984, 'rewards/near_duplicate_penalty/std': 0.02611256204545498, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9807692170143127, 'rewards/distinct_2/std': 0.054392825812101364, 'rewards/total_composite/mean': 0.45705297589302063, 'rewards/total_composite/std': 0.29990771412849426, 'reward': 0.45705297589302063, 'reward_std': 0.2999076843261719, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17957256734371185, 'sampling/sampling_logp_difference/max': 1.7710824012756348, 'sampling/importance_sampling_ratio/min': 0.17014873027801514, 'sampling/importance_sampling_ratio/mean': 1.0418623685836792, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7224462032318115, 'clip_ratio/low_mean': 0.10201781429350376, 'clip_ratio/low_min': 0.10201781429350376, 'clip_ratio/high_mean': 0.03315190505236387, 'clip_ratio/high_max': 0.03315190505236387, 'clip_ratio/region_mean': 0.13516971934586763, 'reward_total_mean': 0.45705297589302063, 'reward_meter_mean': 0.7285923957824707, 'reward_meter_std': 0.3594067692756653, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9673515558242798, 'reward_lexical_plausibility_std': 0.09234381467103958, 'reward_repeat_penalty_mean': 0.9713559150695801, 'reward_repeat_penalty_std': 0.07632871717214584, 'reward_near_duplicate_penalty_mean': 0.9891414642333984, 'reward_near_duplicate_penalty_std': 0.02611256204545498, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9807692170143127, 'reward_distinct_2_std': 0.054392825812101364, 'reward_judge_quality_mean': 0.6262500286102295, 'reward_judge_quality_std': 0.2432481348514557, 'reward_total_composite_mean': 0.45705297589302063, 'reward_total_composite_std': 0.29990771412849426, 'epoch': 0.0}

  4%|▎         | 118/3300 [25:38<8:46:46,  9.93s/it]INFO 04-12 11:11:50 [block_pool.py:378] Successfully reset prefix cache

  4%|▎         | 119/3300 [25:46<8:21:41,  9.46s/it]2026-04-12 11:11:58,947 | INFO | train_grpo_train | metrics_logged mode=train step=119

                                                    
{'loss': 0.0434, 'grad_norm': 23.97408103942871, 'learning_rate': 9.642424242424243e-06, 'num_tokens': 254770.0, 'completions/mean_length': 22.375, 'completions/min_length': 17.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.375, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.7641736268997192, 'rewards/meter/std': 0.37070146203041077, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7512500286102295, 'rewards/judge_quality/std': 0.24976776540279388, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5957629680633545, 'rewards/total_composite/std': 0.3396039307117462, 'reward': 0.5957629680633545, 'reward_std': 0.3396039307117462, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22697778046131134, 'sampling/sampling_logp_difference/max': 1.9863262176513672, 'sampling/importance_sampling_ratio/min': 0.1371985375881195, 'sampling/importance_sampling_ratio/mean': 1.0331143140792847, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.084925338625908, 'clip_ratio/low_mean': 0.07491859421133995, 'clip_ratio/low_min': 0.07491859421133995, 'clip_ratio/high_mean': 0.12119132094085217, 'clip_ratio/high_max': 0.12119132094085217, 'clip_ratio/region_mean': 0.19610991515219212, 'reward_total_mean': 0.5957629680633545, 'reward_meter_mean': 0.7641736268997192, 'reward_meter_std': 0.37070146203041077, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7512500286102295, 'reward_judge_quality_std': 0.24976776540279388, 'reward_total_composite_mean': 0.5957629680633545, 'reward_total_composite_std': 0.3396039307117462, 'epoch': 0.0}

  4%|▎         | 119/3300 [25:46<8:21:41,  9.46s/it]INFO 04-12 11:11:59 [block_pool.py:378] Successfully reset prefix cache

  4%|▎         | 120/3300 [25:56<8:17:28,  9.39s/it]2026-04-12 11:12:08,148 | INFO | train_grpo_train | metrics_logged mode=train step=120

                                                    
{'loss': 0.0807, 'grad_norm': 20.732107162475586, 'learning_rate': 9.63939393939394e-06, 'num_tokens': 256473.0, 'completions/mean_length': 38.875, 'completions/min_length': 30.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.875, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.6519851684570312, 'rewards/meter/std': 0.4235266149044037, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9898897409439087, 'rewards/lexical_plausibility/std': 0.028596237301826477, 'rewards/judge_quality/mean': 0.6299999952316284, 'rewards/judge_quality/std': 0.23071318864822388, 'rewards/repeat_penalty/mean': 0.9855146408081055, 'rewards/repeat_penalty/std': 0.03465944901108742, 'rewards/near_duplicate_penalty/mean': 0.9920557737350464, 'rewards/near_duplicate_penalty/std': 0.01630682684481144, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9931318759918213, 'rewards/distinct_2/std': 0.01942601054906845, 'rewards/total_composite/mean': 0.4616625905036926, 'rewards/total_composite/std': 0.3632766306400299, 'reward': 0.4616625905036926, 'reward_std': 0.3632766306400299, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20872028172016144, 'sampling/sampling_logp_difference/max': 1.4982450008392334, 'sampling/importance_sampling_ratio/min': 0.2235221117734909, 'sampling/importance_sampling_ratio/mean': 1.0724166631698608, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.009061813354492, 'clip_ratio/low_mean': 0.06882300227880478, 'clip_ratio/low_min': 0.06882300227880478, 'clip_ratio/high_mean': 0.06803228054195642, 'clip_ratio/high_max': 0.06803228054195642, 'clip_ratio/region_mean': 0.1368552828207612, 'reward_total_mean': 0.4616625905036926, 'reward_meter_mean': 0.6519851684570312, 'reward_meter_std': 0.4235266149044037, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9898897409439087, 'reward_lexical_plausibility_std': 0.028596237301826477, 'reward_repeat_penalty_mean': 0.9855146408081055, 'reward_repeat_penalty_std': 0.03465944901108742, 'reward_near_duplicate_penalty_mean': 0.9920557737350464, 'reward_near_duplicate_penalty_std': 0.01630682684481144, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9931318759918213, 'reward_distinct_2_std': 0.01942601054906845, 'reward_judge_quality_mean': 0.6299999952316284, 'reward_judge_quality_std': 0.23071318864822388, 'reward_total_composite_mean': 0.4616625905036926, 'reward_total_composite_std': 0.3632766306400299, 'epoch': 0.0}

  4%|▎         | 120/3300 [25:56<8:17:28,  9.39s/it]INFO 04-12 11:12:08 [block_pool.py:378] Successfully reset prefix cache

  4%|▎         | 121/3300 [26:10<9:31:58, 10.80s/it]2026-04-12 11:12:22,230 | INFO | train_grpo_train | metrics_logged mode=train step=121

                                                    
{'loss': 0.1101, 'grad_norm': 4.198246479034424, 'learning_rate': 9.636363636363638e-06, 'num_tokens': 261019.0, 'completions/mean_length': 380.25, 'completions/min_length': 289.0, 'completions/max_length': 470.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 380.25, 'completions/min_terminated_length': 289.0, 'completions/max_terminated_length': 470.0, 'rewards/meter/mean': 0.4872194230556488, 'rewards/meter/std': 0.2961656451225281, 'rewards/count_adherence/mean': 0.9038461446762085, 'rewards/count_adherence/std': 0.05439282953739166, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.9988567233085632, 'rewards/lexical_plausibility/std': 0.0032337172888219357, 'rewards/judge_quality/mean': 0.20000000298023224, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.9843146204948425, 'rewards/repeat_penalty/std': 0.011409228667616844, 'rewards/near_duplicate_penalty/mean': 0.9946587085723877, 'rewards/near_duplicate_penalty/std': 0.002646322827786207, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9895939826965332, 'rewards/distinct_2/std': 0.01048082485795021, 'rewards/total_composite/mean': 0.059502169489860535, 'rewards/total_composite/std': 0.08205050230026245, 'reward': 0.059502169489860535, 'reward_std': 0.08205050230026245, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21086423099040985, 'sampling/sampling_logp_difference/max': 1.8623857498168945, 'sampling/importance_sampling_ratio/min': 0.15530167520046234, 'sampling/importance_sampling_ratio/mean': 1.0560636520385742, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.2698571979999542, 'clip_ratio/low_mean': 0.0990350991487503, 'clip_ratio/low_min': 0.0990350991487503, 'clip_ratio/high_mean': 0.07289245538413525, 'clip_ratio/high_max': 0.07289245538413525, 'clip_ratio/region_mean': 0.17192755453288555, 'reward_total_mean': 0.059502169489860535, 'reward_meter_mean': 0.4872194230556488, 'reward_meter_std': 0.2961656451225281, 'reward_count_adherence_mean': 0.9038461446762085, 'reward_count_adherence_std': 0.05439282953739166, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.9988567233085632, 'reward_lexical_plausibility_std': 0.0032337172888219357, 'reward_repeat_penalty_mean': 0.9843146204948425, 'reward_repeat_penalty_std': 0.011409228667616844, 'reward_near_duplicate_penalty_mean': 0.9946587085723877, 'reward_near_duplicate_penalty_std': 0.002646322827786207, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9895939826965332, 'reward_distinct_2_std': 0.01048082485795021, 'reward_judge_quality_mean': 0.20000000298023224, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.059502169489860535, 'reward_total_composite_std': 0.08205050230026245, 'epoch': 0.0}

  4%|▎         | 121/3300 [26:10<9:31:58, 10.80s/it]INFO 04-12 11:12:22 [block_pool.py:378] Successfully reset prefix cache

  4%|▎         | 122/3300 [26:18<8:53:55, 10.08s/it]2026-04-12 11:12:30,657 | INFO | train_grpo_train | metrics_logged mode=train step=122

                                                    
{'loss': 0.0546, 'grad_norm': 15.544112205505371, 'learning_rate': 9.633333333333335e-06, 'num_tokens': 262466.0, 'completions/mean_length': 44.875, 'completions/min_length': 30.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.875, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.4326896369457245, 'rewards/meter/std': 0.4062103033065796, '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/lexical_plausibility/mean': 0.9808238744735718, 'rewards/lexical_plausibility/std': 0.03631000593304634, 'rewards/judge_quality/mean': 0.38750001788139343, 'rewards/judge_quality/std': 0.3283182382583618, 'rewards/repeat_penalty/mean': 0.9809596538543701, 'rewards/repeat_penalty/std': 0.039783187210559845, 'rewards/near_duplicate_penalty/mean': 0.9892747402191162, 'rewards/near_duplicate_penalty/std': 0.017515402287244797, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9912587404251099, 'rewards/distinct_2/std': 0.024724015966057777, 'rewards/total_composite/mean': 0.18216806650161743, 'rewards/total_composite/std': 0.3033214807510376, 'reward': 0.18216806650161743, 'reward_std': 0.3033214807510376, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.26005348563194275, 'sampling/sampling_logp_difference/max': 1.4416942596435547, 'sampling/importance_sampling_ratio/min': 0.23652668297290802, 'sampling/importance_sampling_ratio/mean': 1.0260099172592163, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.8264652639627457, 'clip_ratio/low_mean': 0.1550660226494074, 'clip_ratio/low_min': 0.1550660226494074, 'clip_ratio/high_mean': 0.04622093215584755, 'clip_ratio/high_max': 0.04622093215584755, 'clip_ratio/region_mean': 0.20128695480525494, 'reward_total_mean': 0.18216806650161743, 'reward_meter_mean': 0.4326896369457245, 'reward_meter_std': 0.4062103033065796, '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_lexical_plausibility_mean': 0.9808238744735718, 'reward_lexical_plausibility_std': 0.03631000593304634, 'reward_repeat_penalty_mean': 0.9809596538543701, 'reward_repeat_penalty_std': 0.039783187210559845, 'reward_near_duplicate_penalty_mean': 0.9892747402191162, 'reward_near_duplicate_penalty_std': 0.017515402287244797, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9912587404251099, 'reward_distinct_2_std': 0.024724015966057777, 'reward_judge_quality_mean': 0.38750001788139343, 'reward_judge_quality_std': 0.3283182382583618, 'reward_total_composite_mean': 0.18216806650161743, 'reward_total_composite_std': 0.3033214807510376, 'epoch': 0.0}

  4%|▎         | 122/3300 [26:18<8:53:55, 10.08s/it]INFO 04-12 11:12:30 [block_pool.py:378] Successfully reset prefix cache

  4%|▎         | 123/3300 [26:27<8:42:19,  9.86s/it]2026-04-12 11:12:40,004 | INFO | train_grpo_train | metrics_logged mode=train step=123

                                                    
{'loss': 0.1375, 'grad_norm': 14.678629875183105, 'learning_rate': 9.63030303030303e-06, 'num_tokens': 264128.0, 'completions/mean_length': 46.75, 'completions/min_length': 33.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.75, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.32004526257514954, 'rewards/meter/std': 0.4370953440666199, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.48374998569488525, 'rewards/judge_quality/std': 0.18196842074394226, 'rewards/repeat_penalty/mean': 0.9907463788986206, 'rewards/repeat_penalty/std': 0.011995292268693447, 'rewards/near_duplicate_penalty/mean': 0.9907463788986206, 'rewards/near_duplicate_penalty/std': 0.011995292268693447, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.13091403245925903, 'rewards/total_composite/std': 0.17915959656238556, 'reward': 0.13091403245925903, 'reward_std': 0.17915959656238556, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21963627636432648, 'sampling/sampling_logp_difference/max': 1.3811006546020508, 'sampling/importance_sampling_ratio/min': 0.2513018250465393, 'sampling/importance_sampling_ratio/mean': 1.0610466003417969, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8643734902143478, 'clip_ratio/low_mean': 0.1139213452115655, 'clip_ratio/low_min': 0.1139213452115655, 'clip_ratio/high_mean': 0.06875000149011612, 'clip_ratio/high_max': 0.06875000149011612, 'clip_ratio/region_mean': 0.18267134670168161, 'reward_total_mean': 0.13091403245925903, 'reward_meter_mean': 0.32004526257514954, 'reward_meter_std': 0.4370953440666199, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9907463788986206, 'reward_repeat_penalty_std': 0.011995292268693447, 'reward_near_duplicate_penalty_mean': 0.9907463788986206, 'reward_near_duplicate_penalty_std': 0.011995292268693447, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.48374998569488525, 'reward_judge_quality_std': 0.18196842074394226, 'reward_total_composite_mean': 0.13091403245925903, 'reward_total_composite_std': 0.17915959656238556, 'epoch': 0.0}

  4%|▎         | 123/3300 [26:27<8:42:19,  9.86s/it]INFO 04-12 11:12:40 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 124/3300 [26:39<9:01:16, 10.23s/it]2026-04-12 11:12:51,075 | INFO | train_grpo_train | metrics_logged mode=train step=124

                                                    
{'loss': 0.071, 'grad_norm': 8.55613899230957, 'learning_rate': 9.627272727272728e-06, 'num_tokens': 266931.0, 'completions/mean_length': 161.375, 'completions/min_length': 135.0, 'completions/max_length': 187.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 161.375, 'completions/min_terminated_length': 135.0, 'completions/max_terminated_length': 187.0, 'rewards/meter/mean': 0.7234280109405518, 'rewards/meter/std': 0.28849291801452637, 'rewards/count_adherence/mean': 0.8571428656578064, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.08864051848649979, 'rewards/repeat_penalty/mean': 0.9556269645690918, 'rewards/repeat_penalty/std': 0.024927888065576553, 'rewards/near_duplicate_penalty/mean': 0.9893919229507446, 'rewards/near_duplicate_penalty/std': 0.004175576847046614, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9658507108688354, 'rewards/distinct_2/std': 0.023846343159675598, 'rewards/total_composite/mean': 0.19653823971748352, 'rewards/total_composite/std': 0.10214091092348099, 'reward': 0.19653823971748352, 'reward_std': 0.10214090347290039, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2302684783935547, 'sampling/sampling_logp_difference/max': 1.5414671897888184, 'sampling/importance_sampling_ratio/min': 0.2203531712293625, 'sampling/importance_sampling_ratio/mean': 1.0580713748931885, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.2126179933547974, 'clip_ratio/low_mean': 0.06961525604128838, 'clip_ratio/low_min': 0.06961525604128838, 'clip_ratio/high_mean': 0.12272296473383904, 'clip_ratio/high_max': 0.12272296473383904, 'clip_ratio/region_mean': 0.1923382207751274, 'reward_total_mean': 0.19653823971748352, 'reward_meter_mean': 0.7234280109405518, 'reward_meter_std': 0.28849291801452637, 'reward_count_adherence_mean': 0.8571428656578064, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9556269645690918, 'reward_repeat_penalty_std': 0.024927888065576553, 'reward_near_duplicate_penalty_mean': 0.9893919229507446, 'reward_near_duplicate_penalty_std': 0.004175576847046614, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9658507108688354, 'reward_distinct_2_std': 0.023846343159675598, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.08864051848649979, 'reward_total_composite_mean': 0.19653823971748352, 'reward_total_composite_std': 0.10214091092348099, 'epoch': 0.0}

  4%|▍         | 124/3300 [26:39<9:01:16, 10.23s/it]INFO 04-12 11:12:51 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 125/3300 [26:48<8:41:56,  9.86s/it]2026-04-12 11:13:00,091 | INFO | train_grpo_train | metrics_logged mode=train step=125

                                                    
{'loss': -0.0016, 'grad_norm': 12.818426132202148, 'learning_rate': 9.624242424242425e-06, 'num_tokens': 268778.0, 'completions/mean_length': 57.875, 'completions/min_length': 43.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.875, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.7629757523536682, 'rewards/meter/std': 0.36042335629463196, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9707098007202148, 'rewards/lexical_plausibility/std': 0.0828452855348587, 'rewards/judge_quality/mean': 0.4625000059604645, 'rewards/judge_quality/std': 0.2938780188560486, 'rewards/repeat_penalty/mean': 0.9815762639045715, 'rewards/repeat_penalty/std': 0.02969047613441944, 'rewards/near_duplicate_penalty/mean': 0.9905714988708496, 'rewards/near_duplicate_penalty/std': 0.007983755320310593, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9908424615859985, 'rewards/distinct_2/std': 0.025901345536112785, 'rewards/total_composite/mean': 0.33154475688934326, 'rewards/total_composite/std': 0.2759409248828888, 'reward': 0.33154475688934326, 'reward_std': 0.2759409248828888, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19727781414985657, 'sampling/sampling_logp_difference/max': 1.3558845520019531, 'sampling/importance_sampling_ratio/min': 0.2577192187309265, 'sampling/importance_sampling_ratio/mean': 1.0571469068527222, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.360900267958641, 'clip_ratio/low_mean': 0.10369003191590309, 'clip_ratio/low_min': 0.10369003191590309, 'clip_ratio/high_mean': 0.10226017236709595, 'clip_ratio/high_max': 0.10226017236709595, 'clip_ratio/region_mean': 0.20595020428299904, 'reward_total_mean': 0.33154475688934326, 'reward_meter_mean': 0.7629757523536682, 'reward_meter_std': 0.36042335629463196, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9707098007202148, 'reward_lexical_plausibility_std': 0.0828452855348587, 'reward_repeat_penalty_mean': 0.9815762639045715, 'reward_repeat_penalty_std': 0.02969047613441944, 'reward_near_duplicate_penalty_mean': 0.9905714988708496, 'reward_near_duplicate_penalty_std': 0.007983755320310593, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9908424615859985, 'reward_distinct_2_std': 0.025901345536112785, 'reward_judge_quality_mean': 0.4625000059604645, 'reward_judge_quality_std': 0.2938780188560486, 'reward_total_composite_mean': 0.33154475688934326, 'reward_total_composite_std': 0.2759409248828888, 'epoch': 0.01}

  4%|▍         | 125/3300 [26:48<8:41:56,  9.86s/it]INFO 04-12 11:13:00 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 126/3300 [26:56<8:18:52,  9.43s/it]2026-04-12 11:13:08,518 | INFO | train_grpo_train | metrics_logged mode=train step=126

                                                    
{'loss': 0.032, 'grad_norm': 21.17633819580078, 'learning_rate': 9.621212121212122e-06, 'num_tokens': 270279.0, 'completions/mean_length': 28.625, 'completions/min_length': 26.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.625, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.6716091632843018, 'rewards/meter/std': 0.3910526633262634, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9916572570800781, 'rewards/lexical_plausibility/std': 0.023596882820129395, 'rewards/judge_quality/mean': 0.6600000262260437, 'rewards/judge_quality/std': 0.2805097699165344, 'rewards/repeat_penalty/mean': 0.9878971576690674, 'rewards/repeat_penalty/std': 0.02144477888941765, 'rewards/near_duplicate_penalty/mean': 0.9878971576690674, 'rewards/near_duplicate_penalty/std': 0.02144477888941765, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.43078815937042236, 'rewards/total_composite/std': 0.3094470798969269, 'reward': 0.43078815937042236, 'reward_std': 0.3094470798969269, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18919698894023895, 'sampling/sampling_logp_difference/max': 1.7103853225708008, 'sampling/importance_sampling_ratio/min': 0.18079611659049988, 'sampling/importance_sampling_ratio/mean': 1.008143424987793, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9981854781508446, 'clip_ratio/low_mean': 0.09711630642414093, 'clip_ratio/low_min': 0.09711630642414093, 'clip_ratio/high_mean': 0.056547620333731174, 'clip_ratio/high_max': 0.056547620333731174, 'clip_ratio/region_mean': 0.1536639267578721, 'reward_total_mean': 0.43078815937042236, 'reward_meter_mean': 0.6716091632843018, 'reward_meter_std': 0.3910526633262634, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9916572570800781, 'reward_lexical_plausibility_std': 0.023596882820129395, 'reward_repeat_penalty_mean': 0.9878971576690674, 'reward_repeat_penalty_std': 0.02144477888941765, 'reward_near_duplicate_penalty_mean': 0.9878971576690674, 'reward_near_duplicate_penalty_std': 0.02144477888941765, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6600000262260437, 'reward_judge_quality_std': 0.2805097699165344, 'reward_total_composite_mean': 0.43078815937042236, 'reward_total_composite_std': 0.3094470798969269, 'epoch': 0.01}

  4%|▍         | 126/3300 [26:56<8:18:52,  9.43s/it]INFO 04-12 11:13:08 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 127/3300 [27:05<8:16:59,  9.40s/it]2026-04-12 11:13:17,835 | INFO | train_grpo_train | metrics_logged mode=train step=127

                                                    
{'loss': -0.0924, 'grad_norm': 13.073019981384277, 'learning_rate': 9.61818181818182e-06, 'num_tokens': 272030.0, 'completions/mean_length': 62.875, 'completions/min_length': 45.0, 'completions/max_length': 88.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.875, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 88.0, 'rewards/meter/mean': 0.3191009759902954, 'rewards/meter/std': 0.29537802934646606, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.0353553369641304, 'rewards/repeat_penalty/mean': 0.9776274561882019, 'rewards/repeat_penalty/std': 0.034937307238578796, 'rewards/near_duplicate_penalty/mean': 0.9852867126464844, 'rewards/near_duplicate_penalty/std': 0.015377524308860302, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9919871687889099, 'rewards/distinct_2/std': 0.02266368828713894, 'rewards/total_composite/mean': 0.12350160628557205, 'rewards/total_composite/std': 0.1331329196691513, 'reward': 0.12350160628557205, 'reward_std': 0.1331329047679901, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2104559987783432, 'sampling/sampling_logp_difference/max': 1.8942790031433105, 'sampling/importance_sampling_ratio/min': 0.1504267454147339, 'sampling/importance_sampling_ratio/mean': 1.0316500663757324, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1464562863111496, 'clip_ratio/low_mean': 0.13627557829022408, 'clip_ratio/low_min': 0.13627557829022408, 'clip_ratio/high_mean': 0.0845691878348589, 'clip_ratio/high_max': 0.0845691878348589, 'clip_ratio/region_mean': 0.22084476612508297, 'reward_total_mean': 0.12350160628557205, 'reward_meter_mean': 0.3191009759902954, 'reward_meter_std': 0.29537802934646606, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9776274561882019, 'reward_repeat_penalty_std': 0.034937307238578796, 'reward_near_duplicate_penalty_mean': 0.9852867126464844, 'reward_near_duplicate_penalty_std': 0.015377524308860302, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9919871687889099, 'reward_distinct_2_std': 0.02266368828713894, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.0353553369641304, 'reward_total_composite_mean': 0.12350160628557205, 'reward_total_composite_std': 0.1331329196691513, 'epoch': 0.01}

  4%|▍         | 127/3300 [27:05<8:16:59,  9.40s/it]INFO 04-12 11:13:18 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 128/3300 [27:15<8:22:32,  9.51s/it]2026-04-12 11:13:27,609 | INFO | train_grpo_train | metrics_logged mode=train step=128

                                                    
{'loss': 0.1126, 'grad_norm': 17.821298599243164, 'learning_rate': 9.615151515151517e-06, 'num_tokens': 273676.0, 'completions/mean_length': 42.75, 'completions/min_length': 32.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.75, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.31968021392822266, 'rewards/meter/std': 0.2793877124786377, '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/lexical_plausibility/mean': 0.9928355813026428, 'rewards/lexical_plausibility/std': 0.020263994112610817, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.10690449178218842, 'rewards/repeat_penalty/mean': 0.9920704364776611, 'rewards/repeat_penalty/std': 0.01435357891023159, 'rewards/near_duplicate_penalty/mean': 0.9920704364776611, 'rewards/near_duplicate_penalty/std': 0.01435357891023159, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.13533230125904083, 'rewards/total_composite/std': 0.11568300426006317, 'reward': 0.13533230125904083, 'reward_std': 0.11568299680948257, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23443779349327087, 'sampling/sampling_logp_difference/max': 1.3514337539672852, 'sampling/importance_sampling_ratio/min': 0.25886884331703186, 'sampling/importance_sampling_ratio/mean': 1.0290594100952148, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.5696670711040497, 'clip_ratio/low_mean': 0.08619480999186635, 'clip_ratio/low_min': 0.08619480999186635, 'clip_ratio/high_mean': 0.13277778029441833, 'clip_ratio/high_max': 0.13277778029441833, 'clip_ratio/region_mean': 0.21897259028628469, 'reward_total_mean': 0.13533230125904083, 'reward_meter_mean': 0.31968021392822266, 'reward_meter_std': 0.2793877124786377, '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_lexical_plausibility_mean': 0.9928355813026428, 'reward_lexical_plausibility_std': 0.020263994112610817, 'reward_repeat_penalty_mean': 0.9920704364776611, 'reward_repeat_penalty_std': 0.01435357891023159, 'reward_near_duplicate_penalty_mean': 0.9920704364776611, 'reward_near_duplicate_penalty_std': 0.01435357891023159, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.10690449178218842, 'reward_total_composite_mean': 0.13533230125904083, 'reward_total_composite_std': 0.11568300426006317, 'epoch': 0.01}

  4%|▍         | 128/3300 [27:15<8:22:32,  9.51s/it]INFO 04-12 11:13:27 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 129/3300 [27:26<8:43:37,  9.91s/it]2026-04-12 11:13:38,435 | INFO | train_grpo_train | metrics_logged mode=train step=129

                                                    
{'loss': -0.0287, 'grad_norm': 11.1714448928833, 'learning_rate': 9.612121212121212e-06, 'num_tokens': 276217.0, 'completions/mean_length': 115.625, 'completions/min_length': 70.0, 'completions/max_length': 179.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 115.625, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 179.0, 'rewards/meter/mean': 0.30494463443756104, 'rewards/meter/std': 0.26648348569869995, 'rewards/count_adherence/mean': 0.9249999523162842, 'rewards/count_adherence/std': 0.1035098284482956, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9899933338165283, 'rewards/lexical_plausibility/std': 0.02830304391682148, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.9876667261123657, 'rewards/repeat_penalty/std': 0.020608430728316307, 'rewards/near_duplicate_penalty/mean': 0.9933052062988281, 'rewards/near_duplicate_penalty/std': 0.0052453503012657166, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9942594766616821, 'rewards/distinct_2/std': 0.016236674040555954, 'rewards/total_composite/mean': 0.06439289450645447, 'rewards/total_composite/std': 0.06670676171779633, 'reward': 0.06439289450645447, 'reward_std': 0.06670676171779633, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2419767826795578, 'sampling/sampling_logp_difference/max': 1.5326776504516602, 'sampling/importance_sampling_ratio/min': 0.2159566432237625, 'sampling/importance_sampling_ratio/mean': 1.0452473163604736, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6886102110147476, 'clip_ratio/low_mean': 0.129961758852005, 'clip_ratio/low_min': 0.129961758852005, 'clip_ratio/high_mean': 0.08939770795404911, 'clip_ratio/high_max': 0.08939770795404911, 'clip_ratio/region_mean': 0.21935946680605412, 'reward_total_mean': 0.06439289450645447, 'reward_meter_mean': 0.30494463443756104, 'reward_meter_std': 0.26648348569869995, 'reward_count_adherence_mean': 0.9249999523162842, 'reward_count_adherence_std': 0.1035098284482956, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9899933338165283, 'reward_lexical_plausibility_std': 0.02830304391682148, 'reward_repeat_penalty_mean': 0.9876667261123657, 'reward_repeat_penalty_std': 0.020608430728316307, 'reward_near_duplicate_penalty_mean': 0.9933052062988281, 'reward_near_duplicate_penalty_std': 0.0052453503012657166, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9942594766616821, 'reward_distinct_2_std': 0.016236674040555954, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.06439289450645447, 'reward_total_composite_std': 0.06670676171779633, 'epoch': 0.01}

  4%|▍         | 129/3300 [27:26<8:43:37,  9.91s/it]INFO 04-12 11:13:38 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 130/3300 [27:37<9:01:51, 10.26s/it]2026-04-12 11:13:49,505 | INFO | train_grpo_train | metrics_logged mode=train step=130

                                                    
{'loss': 0.1365, 'grad_norm': 8.778731346130371, 'learning_rate': 9.60909090909091e-06, 'num_tokens': 278568.0, 'completions/mean_length': 119.875, 'completions/min_length': 81.0, 'completions/max_length': 179.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 119.875, 'completions/min_terminated_length': 81.0, 'completions/max_terminated_length': 179.0, 'rewards/meter/mean': 0.8424458503723145, 'rewards/meter/std': 0.22091150283813477, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.10690449178218842, 'rewards/repeat_penalty/mean': 0.9579651355743408, 'rewards/repeat_penalty/std': 0.07393673807382584, 'rewards/near_duplicate_penalty/mean': 0.9907664060592651, 'rewards/near_duplicate_penalty/std': 0.013159936293959618, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9661681056022644, 'rewards/distinct_2/std': 0.06276010721921921, 'rewards/total_composite/mean': 0.2576933801174164, 'rewards/total_composite/std': 0.12515640258789062, 'reward': 0.2576933801174164, 'reward_std': 0.12515640258789062, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.232686847448349, 'sampling/sampling_logp_difference/max': 1.626357078552246, 'sampling/importance_sampling_ratio/min': 0.1966446340084076, 'sampling/importance_sampling_ratio/mean': 1.038954496383667, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.205583095550537, 'clip_ratio/low_mean': 0.07093797065317631, 'clip_ratio/low_min': 0.07093797065317631, 'clip_ratio/high_mean': 0.10020424611866474, 'clip_ratio/high_max': 0.10020424611866474, 'clip_ratio/region_mean': 0.17114221677184105, 'reward_total_mean': 0.2576933801174164, 'reward_meter_mean': 0.8424458503723145, 'reward_meter_std': 0.22091150283813477, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9579651355743408, 'reward_repeat_penalty_std': 0.07393673807382584, 'reward_near_duplicate_penalty_mean': 0.9907664060592651, 'reward_near_duplicate_penalty_std': 0.013159936293959618, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9661681056022644, 'reward_distinct_2_std': 0.06276010721921921, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.10690449178218842, 'reward_total_composite_mean': 0.2576933801174164, 'reward_total_composite_std': 0.12515640258789062, 'epoch': 0.01}

  4%|▍         | 130/3300 [27:37<9:01:51, 10.26s/it]INFO 04-12 11:13:49 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 131/3300 [27:51<10:01:12, 11.38s/it]2026-04-12 11:14:03,566 | INFO | train_grpo_train | metrics_logged mode=train step=131

                                                     
{'loss': -0.0782, 'grad_norm': 3.8477623462677, 'learning_rate': 9.606060606060607e-06, 'num_tokens': 282878.0, 'completions/mean_length': 329.75, 'completions/min_length': 52.0, 'completions/max_length': 445.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 329.75, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 445.0, 'rewards/meter/mean': 0.8844295740127563, 'rewards/meter/std': 0.11742876470088959, 'rewards/count_adherence/mean': 0.7291666269302368, 'rewards/count_adherence/std': 0.23464766144752502, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.17500001192092896, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/repeat_penalty/mean': 0.984276294708252, 'rewards/repeat_penalty/std': 0.016946520656347275, 'rewards/near_duplicate_penalty/mean': 0.9957332611083984, 'rewards/near_duplicate_penalty/std': 0.002406543819233775, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9884704351425171, 'rewards/distinct_2/std': 0.015221429988741875, 'rewards/total_composite/mean': 0.11510159820318222, 'rewards/total_composite/std': 0.07259103655815125, 'reward': 0.11510159820318222, 'reward_std': 0.07259103655815125, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21964924037456512, 'sampling/sampling_logp_difference/max': 1.4762115478515625, 'sampling/importance_sampling_ratio/min': 0.22850172221660614, 'sampling/importance_sampling_ratio/mean': 1.0544650554656982, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.6216365098953247, 'clip_ratio/low_mean': 0.11001992411911488, 'clip_ratio/low_min': 0.11001992411911488, 'clip_ratio/high_mean': 0.03885163553059101, 'clip_ratio/high_max': 0.03885163553059101, 'clip_ratio/region_mean': 0.1488715596497059, 'reward_total_mean': 0.11510159820318222, 'reward_meter_mean': 0.8844295740127563, 'reward_meter_std': 0.11742876470088959, 'reward_count_adherence_mean': 0.7291666269302368, 'reward_count_adherence_std': 0.23464766144752502, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.984276294708252, 'reward_repeat_penalty_std': 0.016946520656347275, 'reward_near_duplicate_penalty_mean': 0.9957332611083984, 'reward_near_duplicate_penalty_std': 0.002406543819233775, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9884704351425171, 'reward_distinct_2_std': 0.015221429988741875, 'reward_judge_quality_mean': 0.17500001192092896, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.11510159820318222, 'reward_total_composite_std': 0.07259103655815125, 'epoch': 0.01}

  4%|▍         | 131/3300 [27:51<10:01:12, 11.38s/it]INFO 04-12 11:14:03 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 132/3300 [28:00<9:24:44, 10.70s/it] 2026-04-12 11:14:12,609 | INFO | train_grpo_train | metrics_logged mode=train step=132

                                                    
{'loss': 0.1637, 'grad_norm': 19.3759765625, 'learning_rate': 9.603030303030304e-06, 'num_tokens': 284351.0, 'completions/mean_length': 36.125, 'completions/min_length': 27.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.125, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.3161843419075012, 'rewards/meter/std': 0.2382018268108368, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.48374998569488525, 'rewards/judge_quality/std': 0.2241770625114441, 'rewards/repeat_penalty/mean': 0.9833240509033203, 'rewards/repeat_penalty/std': 0.021121129393577576, 'rewards/near_duplicate_penalty/mean': 0.9833240509033203, 'rewards/near_duplicate_penalty/std': 0.021121129393577576, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.16291549801826477, 'rewards/total_composite/std': 0.162691131234169, 'reward': 0.16291549801826477, 'reward_std': 0.1626911461353302, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20212748646736145, 'sampling/sampling_logp_difference/max': 1.7739019393920898, 'sampling/importance_sampling_ratio/min': 0.16966965794563293, 'sampling/importance_sampling_ratio/mean': 1.0468316078186035, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8667328841984272, 'clip_ratio/low_mean': 0.09415096137672663, 'clip_ratio/low_min': 0.09415096137672663, 'clip_ratio/high_mean': 0.04595364350825548, 'clip_ratio/high_max': 0.04595364350825548, 'clip_ratio/region_mean': 0.1401046048849821, 'reward_total_mean': 0.16291549801826477, 'reward_meter_mean': 0.3161843419075012, 'reward_meter_std': 0.2382018268108368, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9833240509033203, 'reward_repeat_penalty_std': 0.021121129393577576, 'reward_near_duplicate_penalty_mean': 0.9833240509033203, 'reward_near_duplicate_penalty_std': 0.021121129393577576, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.48374998569488525, 'reward_judge_quality_std': 0.2241770625114441, 'reward_total_composite_mean': 0.16291549801826477, 'reward_total_composite_std': 0.162691131234169, 'epoch': 0.01}

  4%|▍         | 132/3300 [28:00<9:24:44, 10.70s/it]INFO 04-12 11:14:12 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 133/3300 [28:09<8:57:13, 10.18s/it]2026-04-12 11:14:21,585 | INFO | train_grpo_train | metrics_logged mode=train step=133

                                                    
{'loss': 0.0998, 'grad_norm': 15.089190483093262, 'learning_rate': 9.600000000000001e-06, 'num_tokens': 286069.0, 'completions/mean_length': 50.75, 'completions/min_length': 28.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.75, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.3334487974643707, 'rewards/meter/std': 0.3898901641368866, '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/lexical_plausibility/mean': 0.9014673233032227, 'rewards/lexical_plausibility/std': 0.13922642171382904, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/repeat_penalty/mean': 0.960153341293335, 'rewards/repeat_penalty/std': 0.0854540467262268, 'rewards/near_duplicate_penalty/mean': 0.991403341293335, 'rewards/near_duplicate_penalty/std': 0.010194990783929825, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.0930812805891037, 'rewards/total_composite/std': 0.10575805604457855, 'reward': 0.0930812805891037, 'reward_std': 0.10575805604457855, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2031748741865158, 'sampling/sampling_logp_difference/max': 1.1954617500305176, 'sampling/importance_sampling_ratio/min': 0.30256420373916626, 'sampling/importance_sampling_ratio/mean': 1.0299148559570312, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.989794209599495, 'clip_ratio/low_mean': 0.1021833922713995, 'clip_ratio/low_min': 0.1021833922713995, 'clip_ratio/high_mean': 0.06888170819729567, 'clip_ratio/high_max': 0.06888170819729567, 'clip_ratio/region_mean': 0.17106510046869516, 'reward_total_mean': 0.0930812805891037, 'reward_meter_mean': 0.3334487974643707, 'reward_meter_std': 0.3898901641368866, '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_lexical_plausibility_mean': 0.9014673233032227, 'reward_lexical_plausibility_std': 0.13922642171382904, 'reward_repeat_penalty_mean': 0.960153341293335, 'reward_repeat_penalty_std': 0.0854540467262268, 'reward_near_duplicate_penalty_mean': 0.991403341293335, 'reward_near_duplicate_penalty_std': 0.010194990783929825, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.0930812805891037, 'reward_total_composite_std': 0.10575805604457855, 'epoch': 0.01}

  4%|▍         | 133/3300 [28:09<8:57:13, 10.18s/it]INFO 04-12 11:14:21 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 134/3300 [28:19<8:58:20, 10.20s/it]2026-04-12 11:14:31,839 | INFO | train_grpo_train | metrics_logged mode=train step=134

                                                    
{'loss': 0.0562, 'grad_norm': 9.034200668334961, 'learning_rate': 9.596969696969699e-06, 'num_tokens': 288485.0, 'completions/mean_length': 124.0, 'completions/min_length': 95.0, 'completions/max_length': 177.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 124.0, 'completions/min_terminated_length': 95.0, 'completions/max_terminated_length': 177.0, 'rewards/meter/mean': 0.3226137161254883, 'rewards/meter/std': 0.266395628452301, 'rewards/count_adherence/mean': 0.9791666269302368, 'rewards/count_adherence/std': 0.0589255727827549, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.05345224589109421, 'rewards/repeat_penalty/mean': 0.9902002811431885, 'rewards/repeat_penalty/std': 0.0050697955302894115, 'rewards/near_duplicate_penalty/mean': 0.9902002811431885, 'rewards/near_duplicate_penalty/std': 0.0050697955302894115, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.08946527540683746, 'rewards/total_composite/std': 0.08898165076971054, 'reward': 0.08946527540683746, 'reward_std': 0.08898164331912994, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24211280047893524, 'sampling/sampling_logp_difference/max': 2.0446956157684326, 'sampling/importance_sampling_ratio/min': 0.12941958010196686, 'sampling/importance_sampling_ratio/mean': 1.0471415519714355, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.588059216737747, 'clip_ratio/low_mean': 0.12168584018945694, 'clip_ratio/low_min': 0.12168584018945694, 'clip_ratio/high_mean': 0.08939998969435692, 'clip_ratio/high_max': 0.08939998969435692, 'clip_ratio/region_mean': 0.21108582988381386, 'reward_total_mean': 0.08946527540683746, 'reward_meter_mean': 0.3226137161254883, 'reward_meter_std': 0.266395628452301, 'reward_count_adherence_mean': 0.9791666269302368, 'reward_count_adherence_std': 0.0589255727827549, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9902002811431885, 'reward_repeat_penalty_std': 0.0050697955302894115, 'reward_near_duplicate_penalty_mean': 0.9902002811431885, 'reward_near_duplicate_penalty_std': 0.0050697955302894115, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.05345224589109421, 'reward_total_composite_mean': 0.08946527540683746, 'reward_total_composite_std': 0.08898165076971054, 'epoch': 0.01}

  4%|▍         | 134/3300 [28:19<8:58:20, 10.20s/it]INFO 04-12 11:14:32 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 135/3300 [28:27<8:22:19,  9.52s/it]2026-04-12 11:14:39,777 | INFO | train_grpo_train | metrics_logged mode=train step=135

                                                    
{'loss': -0.0896, 'grad_norm': 20.581274032592773, 'learning_rate': 9.593939393939394e-06, 'num_tokens': 289871.0, 'completions/mean_length': 20.25, 'completions/min_length': 16.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.25, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.6398148536682129, 'rewards/meter/std': 0.41537731885910034, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9554227590560913, 'rewards/lexical_plausibility/std': 0.10317420959472656, 'rewards/judge_quality/mean': 0.5799999833106995, 'rewards/judge_quality/std': 0.3062678575515747, 'rewards/repeat_penalty/mean': 0.7446338534355164, 'rewards/repeat_penalty/std': 0.015177796594798565, 'rewards/near_duplicate_penalty/mean': 0.9928451180458069, 'rewards/near_duplicate_penalty/std': 0.020237062126398087, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.44227364659309387, 'rewards/total_composite/std': 0.37129056453704834, 'reward': 0.44227364659309387, 'reward_std': 0.37129056453704834, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.25178173184394836, 'sampling/sampling_logp_difference/max': 1.6135683059692383, 'sampling/importance_sampling_ratio/min': 0.199175626039505, 'sampling/importance_sampling_ratio/mean': 1.0048104524612427, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.286774158477783, 'clip_ratio/low_mean': 0.12150456942617893, 'clip_ratio/low_min': 0.12150456942617893, 'clip_ratio/high_mean': 0.06586779560893774, 'clip_ratio/high_max': 0.06586779560893774, 'clip_ratio/region_mean': 0.18737236503511667, 'reward_total_mean': 0.44227364659309387, 'reward_meter_mean': 0.6398148536682129, 'reward_meter_std': 0.41537731885910034, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9554227590560913, 'reward_lexical_plausibility_std': 0.10317420959472656, 'reward_repeat_penalty_mean': 0.7446338534355164, 'reward_repeat_penalty_std': 0.015177796594798565, 'reward_near_duplicate_penalty_mean': 0.9928451180458069, 'reward_near_duplicate_penalty_std': 0.020237062126398087, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5799999833106995, 'reward_judge_quality_std': 0.3062678575515747, 'reward_total_composite_mean': 0.44227364659309387, 'reward_total_composite_std': 0.37129056453704834, 'epoch': 0.01}

  4%|▍         | 135/3300 [28:27<8:22:19,  9.52s/it]INFO 04-12 11:14:40 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 136/3300 [28:38<8:42:22,  9.91s/it]2026-04-12 11:14:50,577 | INFO | train_grpo_train | metrics_logged mode=train step=136

                                                    
{'loss': -0.0265, 'grad_norm': 9.126276969909668, 'learning_rate': 9.590909090909091e-06, 'num_tokens': 292355.0, 'completions/mean_length': 109.5, 'completions/min_length': 84.0, 'completions/max_length': 134.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 109.5, 'completions/min_terminated_length': 84.0, 'completions/max_terminated_length': 134.0, 'rewards/meter/mean': 0.6719751358032227, 'rewards/meter/std': 0.31151083111763, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9214310646057129, 'rewards/lexical_plausibility/std': 0.06633367389440536, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.11649647355079651, 'rewards/repeat_penalty/mean': 0.9427301287651062, 'rewards/repeat_penalty/std': 0.07048720121383667, 'rewards/near_duplicate_penalty/mean': 0.9837743639945984, 'rewards/near_duplicate_penalty/std': 0.013407698832452297, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.9651181697845459, 'rewards/distinct_2/std': 0.055662382394075394, 'rewards/total_composite/mean': 0.18346239626407623, 'rewards/total_composite/std': 0.11258508265018463, 'reward': 0.18346239626407623, 'reward_std': 0.11258507519960403, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23017889261245728, 'sampling/sampling_logp_difference/max': 1.4955120086669922, 'sampling/importance_sampling_ratio/min': 0.22413381934165955, 'sampling/importance_sampling_ratio/mean': 1.0508452653884888, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.91239795088768, 'clip_ratio/low_mean': 0.1406170055270195, 'clip_ratio/low_min': 0.1406170055270195, 'clip_ratio/high_mean': 0.0502953976392746, 'clip_ratio/high_max': 0.0502953976392746, 'clip_ratio/region_mean': 0.1909124031662941, 'reward_total_mean': 0.18346239626407623, 'reward_meter_mean': 0.6719751358032227, 'reward_meter_std': 0.31151083111763, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9214310646057129, 'reward_lexical_plausibility_std': 0.06633367389440536, 'reward_repeat_penalty_mean': 0.9427301287651062, 'reward_repeat_penalty_std': 0.07048720121383667, 'reward_near_duplicate_penalty_mean': 0.9837743639945984, 'reward_near_duplicate_penalty_std': 0.013407698832452297, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.9651181697845459, 'reward_distinct_2_std': 0.055662382394075394, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.11649647355079651, 'reward_total_composite_mean': 0.18346239626407623, 'reward_total_composite_std': 0.11258508265018463, 'epoch': 0.01}

  4%|▍         | 136/3300 [28:38<8:42:22,  9.91s/it]INFO 04-12 11:14:50 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 137/3300 [28:48<8:41:44,  9.90s/it]2026-04-12 11:15:00,452 | INFO | train_grpo_train | metrics_logged mode=train step=137

                                                    
{'loss': -0.024, 'grad_norm': 21.552860260009766, 'learning_rate': 9.587878787878789e-06, 'num_tokens': 293852.0, 'completions/mean_length': 35.125, 'completions/min_length': 29.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.125, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.786868155002594, 'rewards/meter/std': 0.2549407482147217, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9798868894577026, 'rewards/lexical_plausibility/std': 0.04668755084276199, 'rewards/judge_quality/mean': 0.5262500047683716, 'rewards/judge_quality/std': 0.3414648175239563, 'rewards/repeat_penalty/mean': 0.9912117719650269, 'rewards/repeat_penalty/std': 0.010851188562810421, 'rewards/near_duplicate_penalty/mean': 0.9912117719650269, 'rewards/near_duplicate_penalty/std': 0.010851188562810421, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4462380111217499, 'rewards/total_composite/std': 0.3615570068359375, 'reward': 0.4462380111217499, 'reward_std': 0.3615570068359375, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20842508971691132, 'sampling/sampling_logp_difference/max': 1.7045316696166992, 'sampling/importance_sampling_ratio/min': 0.18185754120349884, 'sampling/importance_sampling_ratio/mean': 1.0384540557861328, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7997378706932068, 'clip_ratio/low_mean': 0.14936923701316118, 'clip_ratio/low_min': 0.14936923701316118, 'clip_ratio/high_mean': 0.060762768844142556, 'clip_ratio/high_max': 0.060762768844142556, 'clip_ratio/region_mean': 0.21013200585730374, 'reward_total_mean': 0.4462380111217499, 'reward_meter_mean': 0.786868155002594, 'reward_meter_std': 0.2549407482147217, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9798868894577026, 'reward_lexical_plausibility_std': 0.04668755084276199, 'reward_repeat_penalty_mean': 0.9912117719650269, 'reward_repeat_penalty_std': 0.010851188562810421, 'reward_near_duplicate_penalty_mean': 0.9912117719650269, 'reward_near_duplicate_penalty_std': 0.010851188562810421, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5262500047683716, 'reward_judge_quality_std': 0.3414648175239563, 'reward_total_composite_mean': 0.4462380111217499, 'reward_total_composite_std': 0.3615570068359375, 'epoch': 0.01}

  4%|▍         | 137/3300 [28:48<8:41:44,  9.90s/it]INFO 04-12 11:15:00 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 138/3300 [28:57<8:22:58,  9.54s/it]2026-04-12 11:15:09,174 | INFO | train_grpo_train | metrics_logged mode=train step=138

                                                    
{'loss': 0.1688, 'grad_norm': 18.655946731567383, 'learning_rate': 9.584848484848486e-06, 'num_tokens': 295172.0, 'completions/mean_length': 22.0, 'completions/min_length': 17.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.0, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.6750556230545044, 'rewards/meter/std': 0.41225430369377136, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7312500476837158, 'rewards/judge_quality/std': 0.3589046001434326, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.44987037777900696, 'rewards/total_composite/std': 0.40740644931793213, 'reward': 0.44987037777900696, 'reward_std': 0.40740641951560974, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17174120247364044, 'sampling/sampling_logp_difference/max': 1.9114164113998413, 'sampling/importance_sampling_ratio/min': 0.14787079393863678, 'sampling/importance_sampling_ratio/mean': 1.0470362901687622, 'sampling/importance_sampling_ratio/max': 1.9874305725097656, 'entropy': 1.164416842162609, 'clip_ratio/low_mean': 0.08753293938934803, 'clip_ratio/low_min': 0.08753293938934803, 'clip_ratio/high_mean': 0.05992756085470319, 'clip_ratio/high_max': 0.05992756085470319, 'clip_ratio/region_mean': 0.14746050024405122, 'reward_total_mean': 0.44987037777900696, 'reward_meter_mean': 0.6750556230545044, 'reward_meter_std': 0.41225430369377136, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7312500476837158, 'reward_judge_quality_std': 0.3589046001434326, 'reward_total_composite_mean': 0.44987037777900696, 'reward_total_composite_std': 0.40740644931793213, 'epoch': 0.01}

  4%|▍         | 138/3300 [28:57<8:22:58,  9.54s/it]INFO 04-12 11:15:09 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 139/3300 [29:07<8:28:04,  9.64s/it]2026-04-12 11:15:19,051 | INFO | train_grpo_train | metrics_logged mode=train step=139

                                                    
{'loss': 0.1143, 'grad_norm': 10.605066299438477, 'learning_rate': 9.581818181818181e-06, 'num_tokens': 297230.0, 'completions/mean_length': 87.25, 'completions/min_length': 78.0, 'completions/max_length': 123.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 87.25, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 123.0, 'rewards/meter/mean': 0.7937848567962646, 'rewards/meter/std': 0.2852884829044342, '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/lexical_plausibility/mean': 0.9791418313980103, 'rewards/lexical_plausibility/std': 0.03369014337658882, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.9874218106269836, 'rewards/repeat_penalty/std': 0.0075856903567910194, 'rewards/near_duplicate_penalty/mean': 0.9874218106269836, 'rewards/near_duplicate_penalty/std': 0.0075856903567910194, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.25178635120391846, 'rewards/total_composite/std': 0.11754924803972244, 'reward': 0.25178635120391846, 'reward_std': 0.11754924058914185, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22691038250923157, 'sampling/sampling_logp_difference/max': 1.4515964984893799, 'sampling/importance_sampling_ratio/min': 0.23419609665870667, 'sampling/importance_sampling_ratio/mean': 1.0392982959747314, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6376554667949677, 'clip_ratio/low_mean': 0.10384871438145638, 'clip_ratio/low_min': 0.10384871438145638, 'clip_ratio/high_mean': 0.11633804067969322, 'clip_ratio/high_max': 0.11633804067969322, 'clip_ratio/region_mean': 0.2201867550611496, 'reward_total_mean': 0.25178635120391846, 'reward_meter_mean': 0.7937848567962646, 'reward_meter_std': 0.2852884829044342, '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_lexical_plausibility_mean': 0.9791418313980103, 'reward_lexical_plausibility_std': 0.03369014337658882, 'reward_repeat_penalty_mean': 0.9874218106269836, 'reward_repeat_penalty_std': 0.0075856903567910194, 'reward_near_duplicate_penalty_mean': 0.9874218106269836, 'reward_near_duplicate_penalty_std': 0.0075856903567910194, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.25178635120391846, 'reward_total_composite_std': 0.11754924803972244, 'epoch': 0.01}

  4%|▍         | 139/3300 [29:07<8:28:04,  9.64s/it]INFO 04-12 11:15:19 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 140/3300 [29:17<8:45:25,  9.98s/it]2026-04-12 11:15:29,803 | INFO | train_grpo_train | metrics_logged mode=train step=140

                                                    
{'loss': 0.1606, 'grad_norm': 11.118247985839844, 'learning_rate': 9.57878787878788e-06, 'num_tokens': 299498.0, 'completions/mean_length': 109.5, 'completions/min_length': 76.0, 'completions/max_length': 145.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 109.5, 'completions/min_terminated_length': 76.0, 'completions/max_terminated_length': 145.0, 'rewards/meter/mean': 0.594388484954834, 'rewards/meter/std': 0.3843231499195099, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.11649646610021591, 'rewards/repeat_penalty/mean': 0.982245683670044, 'rewards/repeat_penalty/std': 0.020189322531223297, 'rewards/near_duplicate_penalty/mean': 0.9862920045852661, 'rewards/near_duplicate_penalty/std': 0.010312417522072792, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9958193898200989, 'rewards/distinct_2/std': 0.0118245305493474, 'rewards/total_composite/mean': 0.19535231590270996, 'rewards/total_composite/std': 0.14439724385738373, 'reward': 0.19535231590270996, 'reward_std': 0.14439725875854492, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2408737689256668, 'sampling/sampling_logp_difference/max': 1.7839741706848145, 'sampling/importance_sampling_ratio/min': 0.16796928644180298, 'sampling/importance_sampling_ratio/mean': 1.0467760562896729, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6270385682582855, 'clip_ratio/low_mean': 0.11177855171263218, 'clip_ratio/low_min': 0.11177855171263218, 'clip_ratio/high_mean': 0.09924664907157421, 'clip_ratio/high_max': 0.09924664907157421, 'clip_ratio/region_mean': 0.2110252007842064, 'reward_total_mean': 0.19535231590270996, 'reward_meter_mean': 0.594388484954834, 'reward_meter_std': 0.3843231499195099, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.982245683670044, 'reward_repeat_penalty_std': 0.020189322531223297, 'reward_near_duplicate_penalty_mean': 0.9862920045852661, 'reward_near_duplicate_penalty_std': 0.010312417522072792, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9958193898200989, 'reward_distinct_2_std': 0.0118245305493474, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.11649646610021591, 'reward_total_composite_mean': 0.19535231590270996, 'reward_total_composite_std': 0.14439724385738373, 'epoch': 0.01}

  4%|▍         | 140/3300 [29:17<8:45:25,  9.98s/it]INFO 04-12 11:15:30 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 141/3300 [29:26<8:20:09,  9.50s/it]2026-04-12 11:15:38,190 | INFO | train_grpo_train | metrics_logged mode=train step=141

                                                    
{'loss': 0.0318, 'grad_norm': 23.561901092529297, 'learning_rate': 9.575757575757576e-06, 'num_tokens': 300958.0, 'completions/mean_length': 27.5, 'completions/min_length': 23.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.5, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.4667555093765259, 'rewards/meter/std': 0.39842984080314636, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.5299999713897705, 'rewards/judge_quality/std': 0.24512389302253723, 'rewards/repeat_penalty/mean': 0.9685021638870239, 'rewards/repeat_penalty/std': 0.04899715259671211, 'rewards/near_duplicate_penalty/mean': 0.981907844543457, 'rewards/near_duplicate_penalty/std': 0.027928875759243965, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9862637519836426, 'rewards/distinct_2/std': 0.0388520210981369, 'rewards/total_composite/mean': 0.26016759872436523, 'rewards/total_composite/std': 0.23429597914218903, 'reward': 0.26016759872436523, 'reward_std': 0.23429596424102783, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24755442142486572, 'sampling/sampling_logp_difference/max': 1.834871768951416, 'sampling/importance_sampling_ratio/min': 0.15963397920131683, 'sampling/importance_sampling_ratio/mean': 1.042060375213623, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5945236384868622, 'clip_ratio/low_mean': 0.14359890297055244, 'clip_ratio/low_min': 0.14359890297055244, 'clip_ratio/high_mean': 0.15771441347897053, 'clip_ratio/high_max': 0.15771441347897053, 'clip_ratio/region_mean': 0.30131331644952297, 'reward_total_mean': 0.26016759872436523, 'reward_meter_mean': 0.4667555093765259, 'reward_meter_std': 0.39842984080314636, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.9685021638870239, 'reward_repeat_penalty_std': 0.04899715259671211, 'reward_near_duplicate_penalty_mean': 0.981907844543457, 'reward_near_duplicate_penalty_std': 0.027928875759243965, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9862637519836426, 'reward_distinct_2_std': 0.0388520210981369, 'reward_judge_quality_mean': 0.5299999713897705, 'reward_judge_quality_std': 0.24512389302253723, 'reward_total_composite_mean': 0.26016759872436523, 'reward_total_composite_std': 0.23429597914218903, 'epoch': 0.01}

  4%|▍         | 141/3300 [29:26<8:20:09,  9.50s/it]INFO 04-12 11:15:38 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 142/3300 [29:35<8:25:12,  9.60s/it]2026-04-12 11:15:48,019 | INFO | train_grpo_train | metrics_logged mode=train step=142

                                                    
{'loss': 0.159, 'grad_norm': 15.81177043914795, 'learning_rate': 9.572727272727273e-06, 'num_tokens': 302747.0, 'completions/mean_length': 57.625, 'completions/min_length': 50.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.625, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.4928305745124817, 'rewards/meter/std': 0.39109230041503906, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9955357313156128, 'rewards/lexical_plausibility/std': 0.008266246877610683, 'rewards/judge_quality/mean': 0.41249996423721313, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.9902790784835815, 'rewards/repeat_penalty/std': 0.0203789584338665, 'rewards/near_duplicate_penalty/mean': 0.9971174597740173, 'rewards/near_duplicate_penalty/std': 0.004312872886657715, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9931318759918213, 'rewards/distinct_2/std': 0.01942601054906845, 'rewards/total_composite/mean': 0.21449744701385498, 'rewards/total_composite/std': 0.18172529339790344, 'reward': 0.21449744701385498, 'reward_std': 0.18172527849674225, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23863141238689423, 'sampling/sampling_logp_difference/max': 1.5719995498657227, 'sampling/importance_sampling_ratio/min': 0.20762960612773895, 'sampling/importance_sampling_ratio/mean': 1.0410646200180054, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.552950456738472, 'clip_ratio/low_mean': 0.0867729727178812, 'clip_ratio/low_min': 0.0867729727178812, 'clip_ratio/high_mean': 0.11342363432049751, 'clip_ratio/high_max': 0.11342363432049751, 'clip_ratio/region_mean': 0.20019660703837872, 'reward_total_mean': 0.21449744701385498, 'reward_meter_mean': 0.4928305745124817, 'reward_meter_std': 0.39109230041503906, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9955357313156128, 'reward_lexical_plausibility_std': 0.008266246877610683, 'reward_repeat_penalty_mean': 0.9902790784835815, 'reward_repeat_penalty_std': 0.0203789584338665, 'reward_near_duplicate_penalty_mean': 0.9971174597740173, 'reward_near_duplicate_penalty_std': 0.004312872886657715, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9931318759918213, 'reward_distinct_2_std': 0.01942601054906845, 'reward_judge_quality_mean': 0.41249996423721313, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.21449744701385498, 'reward_total_composite_std': 0.18172529339790344, 'epoch': 0.01}

  4%|▍         | 142/3300 [29:35<8:25:12,  9.60s/it]INFO 04-12 11:15:48 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 143/3300 [29:47<8:54:30, 10.16s/it]2026-04-12 11:15:59,484 | INFO | train_grpo_train | metrics_logged mode=train step=143

                                                    
{'loss': 0.071, 'grad_norm': 11.606022834777832, 'learning_rate': 9.56969696969697e-06, 'num_tokens': 305041.0, 'completions/mean_length': 84.75, 'completions/min_length': 69.0, 'completions/max_length': 128.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 84.75, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 128.0, 'rewards/meter/mean': 0.48728397488594055, 'rewards/meter/std': 0.3431295156478882, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5324999690055847, 'rewards/judge_quality/std': 0.21887701749801636, 'rewards/repeat_penalty/mean': 0.9933649301528931, 'rewards/repeat_penalty/std': 0.005677627865225077, 'rewards/near_duplicate_penalty/mean': 0.9933649301528931, 'rewards/near_duplicate_penalty/std': 0.005677627865225077, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.23502235114574432, 'rewards/total_composite/std': 0.14621780812740326, 'reward': 0.23502235114574432, 'reward_std': 0.14621782302856445, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22438229620456696, 'sampling/sampling_logp_difference/max': 1.5458917617797852, 'sampling/importance_sampling_ratio/min': 0.21312174201011658, 'sampling/importance_sampling_ratio/mean': 1.0267876386642456, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.5516919791698456, 'clip_ratio/low_mean': 0.09482662566006184, 'clip_ratio/low_min': 0.09482662566006184, 'clip_ratio/high_mean': 0.10475712828338146, 'clip_ratio/high_max': 0.10475712828338146, 'clip_ratio/region_mean': 0.1995837539434433, 'reward_total_mean': 0.23502235114574432, 'reward_meter_mean': 0.48728397488594055, 'reward_meter_std': 0.3431295156478882, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9933649301528931, 'reward_repeat_penalty_std': 0.005677627865225077, 'reward_near_duplicate_penalty_mean': 0.9933649301528931, 'reward_near_duplicate_penalty_std': 0.005677627865225077, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5324999690055847, 'reward_judge_quality_std': 0.21887701749801636, 'reward_total_composite_mean': 0.23502235114574432, 'reward_total_composite_std': 0.14621780812740326, 'epoch': 0.01}

  4%|▍         | 143/3300 [29:47<8:54:30, 10.16s/it]INFO 04-12 11:15:59 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 144/3300 [29:56<8:37:13,  9.83s/it]2026-04-12 11:16:08,559 | INFO | train_grpo_train | metrics_logged mode=train step=144

                                                    
{'loss': 0.0821, 'grad_norm': 18.464693069458008, 'learning_rate': 9.566666666666668e-06, 'num_tokens': 306651.0, 'completions/mean_length': 43.25, 'completions/min_length': 31.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.25, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.7663955092430115, 'rewards/meter/std': 0.35128188133239746, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9875808954238892, 'rewards/lexical_plausibility/std': 0.035126470029354095, 'rewards/judge_quality/mean': 0.6837499737739563, 'rewards/judge_quality/std': 0.27871838212013245, 'rewards/repeat_penalty/mean': 0.9908764958381653, 'rewards/repeat_penalty/std': 0.014119063504040241, 'rewards/near_duplicate_penalty/mean': 0.9908764958381653, 'rewards/near_duplicate_penalty/std': 0.014119063504040241, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5276665091514587, 'rewards/total_composite/std': 0.31427237391471863, 'reward': 0.5276665091514587, 'reward_std': 0.31427237391471863, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1817430704832077, 'sampling/sampling_logp_difference/max': 1.4361786842346191, 'sampling/importance_sampling_ratio/min': 0.2378348708152771, 'sampling/importance_sampling_ratio/mean': 1.0461962223052979, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7290900349617004, 'clip_ratio/low_mean': 0.09043432958424091, 'clip_ratio/low_min': 0.09043432958424091, 'clip_ratio/high_mean': 0.077738793566823, 'clip_ratio/high_max': 0.077738793566823, 'clip_ratio/region_mean': 0.16817312315106392, 'reward_total_mean': 0.5276665091514587, 'reward_meter_mean': 0.7663955092430115, 'reward_meter_std': 0.35128188133239746, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9875808954238892, 'reward_lexical_plausibility_std': 0.035126470029354095, 'reward_repeat_penalty_mean': 0.9908764958381653, 'reward_repeat_penalty_std': 0.014119063504040241, 'reward_near_duplicate_penalty_mean': 0.9908764958381653, 'reward_near_duplicate_penalty_std': 0.014119063504040241, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6837499737739563, 'reward_judge_quality_std': 0.27871838212013245, 'reward_total_composite_mean': 0.5276665091514587, 'reward_total_composite_std': 0.31427237391471863, 'epoch': 0.01}

  4%|▍         | 144/3300 [29:56<8:37:13,  9.83s/it]INFO 04-12 11:16:08 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 145/3300 [30:07<8:59:00, 10.25s/it]2026-04-12 11:16:19,783 | INFO | train_grpo_train | metrics_logged mode=train step=145

                                                    
{'loss': -0.0265, 'grad_norm': 11.207968711853027, 'learning_rate': 9.563636363636365e-06, 'num_tokens': 308810.0, 'completions/mean_length': 88.875, 'completions/min_length': 70.0, 'completions/max_length': 121.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 88.875, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 121.0, 'rewards/meter/mean': 0.6374506950378418, 'rewards/meter/std': 0.33923736214637756, '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/lexical_plausibility/mean': 0.9803248643875122, 'rewards/lexical_plausibility/std': 0.03233964368700981, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.11649646610021591, 'rewards/repeat_penalty/mean': 0.9557034969329834, 'rewards/repeat_penalty/std': 0.07959376275539398, 'rewards/near_duplicate_penalty/mean': 0.9801782369613647, 'rewards/near_duplicate_penalty/std': 0.027161359786987305, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9737154841423035, 'rewards/distinct_2/std': 0.05863630771636963, 'rewards/total_composite/mean': 0.20957598090171814, 'rewards/total_composite/std': 0.1560916304588318, 'reward': 0.20957598090171814, 'reward_std': 0.15609164535999298, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22422216832637787, 'sampling/sampling_logp_difference/max': 1.4223108291625977, 'sampling/importance_sampling_ratio/min': 0.24115610122680664, 'sampling/importance_sampling_ratio/mean': 1.0311670303344727, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.891635835170746, 'clip_ratio/low_mean': 0.08123719692230225, 'clip_ratio/low_min': 0.08123719692230225, 'clip_ratio/high_mean': 0.09273464418947697, 'clip_ratio/high_max': 0.09273464418947697, 'clip_ratio/region_mean': 0.1739718411117792, 'reward_total_mean': 0.20957598090171814, 'reward_meter_mean': 0.6374506950378418, 'reward_meter_std': 0.33923736214637756, '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_lexical_plausibility_mean': 0.9803248643875122, 'reward_lexical_plausibility_std': 0.03233964368700981, 'reward_repeat_penalty_mean': 0.9557034969329834, 'reward_repeat_penalty_std': 0.07959376275539398, 'reward_near_duplicate_penalty_mean': 0.9801782369613647, 'reward_near_duplicate_penalty_std': 0.027161359786987305, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9737154841423035, 'reward_distinct_2_std': 0.05863630771636963, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.11649646610021591, 'reward_total_composite_mean': 0.20957598090171814, 'reward_total_composite_std': 0.1560916304588318, 'epoch': 0.01}

  4%|▍         | 145/3300 [30:07<8:59:00, 10.25s/it]INFO 04-12 11:16:20 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 146/3300 [30:16<8:40:16,  9.90s/it]2026-04-12 11:16:28,857 | INFO | train_grpo_train | metrics_logged mode=train step=146

                                                    
{'loss': 0.0157, 'grad_norm': 18.393598556518555, 'learning_rate': 9.56060606060606e-06, 'num_tokens': 310422.0, 'completions/mean_length': 38.5, 'completions/min_length': 31.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.5, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.9328004121780396, 'rewards/meter/std': 0.07446634769439697, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9797683954238892, 'rewards/lexical_plausibility/std': 0.03873474895954132, 'rewards/judge_quality/mean': 0.4962500333786011, 'rewards/judge_quality/std': 0.2371821254491806, 'rewards/repeat_penalty/mean': 0.9964973330497742, 'rewards/repeat_penalty/std': 0.0061165690422058105, 'rewards/near_duplicate_penalty/mean': 0.9964973330497742, 'rewards/near_duplicate_penalty/std': 0.0061165690422058105, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4638367295265198, 'rewards/total_composite/std': 0.23023727536201477, 'reward': 0.4638367295265198, 'reward_std': 0.23023729026317596, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23962076008319855, 'sampling/sampling_logp_difference/max': 1.4827651977539062, 'sampling/importance_sampling_ratio/min': 0.2270091027021408, 'sampling/importance_sampling_ratio/mean': 1.0683977603912354, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.1862524151802063, 'clip_ratio/low_mean': 0.1564951315522194, 'clip_ratio/low_min': 0.1564951315522194, 'clip_ratio/high_mean': 0.05373218283057213, 'clip_ratio/high_max': 0.05373218283057213, 'clip_ratio/region_mean': 0.21022731438279152, 'reward_total_mean': 0.4638367295265198, 'reward_meter_mean': 0.9328004121780396, 'reward_meter_std': 0.07446634769439697, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9797683954238892, 'reward_lexical_plausibility_std': 0.03873474895954132, 'reward_repeat_penalty_mean': 0.9964973330497742, 'reward_repeat_penalty_std': 0.0061165690422058105, 'reward_near_duplicate_penalty_mean': 0.9964973330497742, 'reward_near_duplicate_penalty_std': 0.0061165690422058105, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4962500333786011, 'reward_judge_quality_std': 0.2371821254491806, 'reward_total_composite_mean': 0.4638367295265198, 'reward_total_composite_std': 0.23023727536201477, 'epoch': 0.01}

  4%|▍         | 146/3300 [30:16<8:40:16,  9.90s/it]INFO 04-12 11:16:29 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 147/3300 [30:30<9:42:04, 11.08s/it]2026-04-12 11:16:42,683 | INFO | train_grpo_train | metrics_logged mode=train step=147

                                                    
{'loss': 0.0428, 'grad_norm': 4.390659332275391, 'learning_rate': 9.55757575757576e-06, 'num_tokens': 315003.0, 'completions/mean_length': 359.625, 'completions/min_length': 319.0, 'completions/max_length': 421.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 359.625, 'completions/min_terminated_length': 319.0, 'completions/max_terminated_length': 421.0, 'rewards/meter/mean': 0.8243576288223267, 'rewards/meter/std': 0.17478956282138824, 'rewards/count_adherence/mean': 0.9416666626930237, 'rewards/count_adherence/std': 0.06606874614953995, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.9937499761581421, 'rewards/lexical_plausibility/std': 0.01767767407000065, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.9944760799407959, 'rewards/repeat_penalty/std': 0.006702517159283161, 'rewards/near_duplicate_penalty/mean': 0.9976387023925781, 'rewards/near_duplicate_penalty/std': 0.0007780436426401138, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9968271255493164, 'rewards/distinct_2/std': 0.006180260796099901, 'rewards/total_composite/mean': 0.1365898847579956, 'rewards/total_composite/std': 0.1328396201133728, 'reward': 0.1365898847579956, 'reward_std': 0.132839635014534, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23052802681922913, 'sampling/sampling_logp_difference/max': 1.9009475708007812, 'sampling/importance_sampling_ratio/min': 0.14942696690559387, 'sampling/importance_sampling_ratio/mean': 1.0586735010147095, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.92021045088768, 'clip_ratio/low_mean': 0.07059000246226788, 'clip_ratio/low_min': 0.07059000246226788, 'clip_ratio/high_mean': 0.09842309728264809, 'clip_ratio/high_max': 0.09842309728264809, 'clip_ratio/region_mean': 0.16901309974491596, 'reward_total_mean': 0.1365898847579956, 'reward_meter_mean': 0.8243576288223267, 'reward_meter_std': 0.17478956282138824, 'reward_count_adherence_mean': 0.9416666626930237, 'reward_count_adherence_std': 0.06606874614953995, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.9937499761581421, 'reward_lexical_plausibility_std': 0.01767767407000065, 'reward_repeat_penalty_mean': 0.9944760799407959, 'reward_repeat_penalty_std': 0.006702517159283161, 'reward_near_duplicate_penalty_mean': 0.9976387023925781, 'reward_near_duplicate_penalty_std': 0.0007780436426401138, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9968271255493164, 'reward_distinct_2_std': 0.006180260796099901, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.1365898847579956, 'reward_total_composite_std': 0.1328396201133728, 'epoch': 0.01}

  4%|▍         | 147/3300 [30:30<9:42:04, 11.08s/it]INFO 04-12 11:16:42 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 148/3300 [30:40<9:24:01, 10.74s/it]2026-04-12 11:16:52,627 | INFO | train_grpo_train | metrics_logged mode=train step=148

                                                    
{'loss': 0.1382, 'grad_norm': 12.24144458770752, 'learning_rate': 9.554545454545455e-06, 'num_tokens': 317081.0, 'completions/mean_length': 79.75, 'completions/min_length': 65.0, 'completions/max_length': 94.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.75, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.5113064646720886, 'rewards/meter/std': 0.30903738737106323, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9957386255264282, 'rewards/lexical_plausibility/std': 0.012052967213094234, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/repeat_penalty/mean': 0.996456503868103, 'rewards/repeat_penalty/std': 0.004235151689499617, 'rewards/near_duplicate_penalty/mean': 0.996456503868103, 'rewards/near_duplicate_penalty/std': 0.004235151689499617, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.15663936734199524, 'rewards/total_composite/std': 0.11305032670497894, 'reward': 0.15663936734199524, 'reward_std': 0.11305032670497894, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23758473992347717, 'sampling/sampling_logp_difference/max': 2.0828194618225098, 'sampling/importance_sampling_ratio/min': 0.13883207738399506, 'sampling/importance_sampling_ratio/mean': 1.038945198059082, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.9721997678279877, 'clip_ratio/low_mean': 0.11492341384291649, 'clip_ratio/low_min': 0.11492341384291649, 'clip_ratio/high_mean': 0.07696209475398064, 'clip_ratio/high_max': 0.07696209475398064, 'clip_ratio/region_mean': 0.19188550859689713, 'reward_total_mean': 0.15663936734199524, 'reward_meter_mean': 0.5113064646720886, 'reward_meter_std': 0.30903738737106323, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9957386255264282, 'reward_lexical_plausibility_std': 0.012052967213094234, 'reward_repeat_penalty_mean': 0.996456503868103, 'reward_repeat_penalty_std': 0.004235151689499617, 'reward_near_duplicate_penalty_mean': 0.996456503868103, 'reward_near_duplicate_penalty_std': 0.004235151689499617, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.15663936734199524, 'reward_total_composite_std': 0.11305032670497894, 'epoch': 0.01}

  4%|▍         | 148/3300 [30:40<9:24:01, 10.74s/it]INFO 04-12 11:16:52 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 149/3300 [30:49<8:48:20, 10.06s/it]2026-04-12 11:17:01,111 | INFO | train_grpo_train | metrics_logged mode=train step=149

                                                    
{'loss': 0.1713, 'grad_norm': 24.899127960205078, 'learning_rate': 9.551515151515152e-06, 'num_tokens': 318752.0, 'completions/mean_length': 22.875, 'completions/min_length': 16.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.875, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.6699537038803101, 'rewards/meter/std': 0.4266339838504791, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9975490570068359, 'rewards/lexical_plausibility/std': 0.006932419259101152, 'rewards/judge_quality/mean': 0.7637500166893005, 'rewards/judge_quality/std': 0.3218667507171631, 'rewards/repeat_penalty/mean': 0.7194664478302002, 'rewards/repeat_penalty/std': 0.02592424489557743, 'rewards/near_duplicate_penalty/mean': 0.9592885971069336, 'rewards/near_duplicate_penalty/std': 0.03456565737724304, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.572995662689209, 'rewards/total_composite/std': 0.4193941056728363, 'reward': 0.572995662689209, 'reward_std': 0.4193940758705139, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15369386970996857, 'sampling/sampling_logp_difference/max': 1.276634693145752, 'sampling/importance_sampling_ratio/min': 0.2789745330810547, 'sampling/importance_sampling_ratio/mean': 1.0236783027648926, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3106824792921543, 'clip_ratio/low_mean': 0.05043184058740735, 'clip_ratio/low_min': 0.05043184058740735, 'clip_ratio/high_mean': 0.07421875, 'clip_ratio/high_max': 0.07421875, 'clip_ratio/region_mean': 0.12465059058740735, 'reward_total_mean': 0.572995662689209, 'reward_meter_mean': 0.6699537038803101, 'reward_meter_std': 0.4266339838504791, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9975490570068359, 'reward_lexical_plausibility_std': 0.006932419259101152, 'reward_repeat_penalty_mean': 0.7194664478302002, 'reward_repeat_penalty_std': 0.02592424489557743, 'reward_near_duplicate_penalty_mean': 0.9592885971069336, 'reward_near_duplicate_penalty_std': 0.03456565737724304, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7637500166893005, 'reward_judge_quality_std': 0.3218667507171631, 'reward_total_composite_mean': 0.572995662689209, 'reward_total_composite_std': 0.4193941056728363, 'epoch': 0.01}

  5%|▍         | 149/3300 [30:49<8:48:20, 10.06s/it]INFO 04-12 11:17:01 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 150/3300 [30:58<8:38:17,  9.87s/it]2026-04-12 11:17:10,543 | INFO | train_grpo_train | metrics_logged mode=train step=150

                                                    
{'loss': 0.0102, 'grad_norm': 12.77075481414795, 'learning_rate': 9.54848484848485e-06, 'num_tokens': 320410.0, 'completions/mean_length': 49.25, 'completions/min_length': 39.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.25, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.7772743701934814, 'rewards/meter/std': 0.3536442816257477, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.48374998569488525, 'rewards/judge_quality/std': 0.18196842074394226, 'rewards/repeat_penalty/mean': 0.9507493376731873, 'rewards/repeat_penalty/std': 0.13930194079875946, 'rewards/near_duplicate_penalty/mean': 0.990851879119873, 'rewards/near_duplicate_penalty/std': 0.02587468922138214, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9839743375778198, 'rewards/distinct_2/std': 0.045327357947826385, 'rewards/total_composite/mean': 0.3888828456401825, 'rewards/total_composite/std': 0.2589547634124756, 'reward': 0.3888828456401825, 'reward_std': 0.2589547634124756, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20877063274383545, 'sampling/sampling_logp_difference/max': 1.5486278533935547, 'sampling/importance_sampling_ratio/min': 0.2125394195318222, 'sampling/importance_sampling_ratio/mean': 1.0580813884735107, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.122045159339905, 'clip_ratio/low_mean': 0.08299900777637959, 'clip_ratio/low_min': 0.08299900777637959, 'clip_ratio/high_mean': 0.06708143837749958, 'clip_ratio/high_max': 0.06708143837749958, 'clip_ratio/region_mean': 0.15008044615387917, 'reward_total_mean': 0.3888828456401825, 'reward_meter_mean': 0.7772743701934814, 'reward_meter_std': 0.3536442816257477, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9507493376731873, 'reward_repeat_penalty_std': 0.13930194079875946, 'reward_near_duplicate_penalty_mean': 0.990851879119873, 'reward_near_duplicate_penalty_std': 0.02587468922138214, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9839743375778198, 'reward_distinct_2_std': 0.045327357947826385, 'reward_judge_quality_mean': 0.48374998569488525, 'reward_judge_quality_std': 0.18196842074394226, 'reward_total_composite_mean': 0.3888828456401825, 'reward_total_composite_std': 0.2589547634124756, 'epoch': 0.01}

  5%|▍         | 150/3300 [30:58<8:38:17,  9.87s/it]INFO 04-12 11:17:10 [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 [00:10<00:59,  5.38s/it]

 23%|██▎       | 3/13 [00:17<01:01,  6.11s/it]

 31%|███       | 4/13 [00:27<01:05,  7.33s/it]

 38%|███▊      | 5/13 [00:38<01:10,  8.76s/it]

 46%|████▌     | 6/13 [00:49<01:06,  9.50s/it]

 54%|█████▍    | 7/13 [00:58<00:56,  9.43s/it]

 62%|██████▏   | 8/13 [01:11<00:51, 10.31s/it]

 69%|██████▉   | 9/13 [01:23<00:44, 11.01s/it]

 77%|███████▋  | 10/13 [01:31<00:30, 10.15s/it]

 85%|████████▍ | 11/13 [01:42<00:20, 10.26s/it]

 92%|█████████▏| 12/13 [01:54<00:10, 10.73s/it]

100%|██████████| 13/13 [02:05<00:00, 10.79s/it]2026-04-12 11:19:26,428 | INFO | train_grpo_train | metrics_logged mode=eval step=150

                                                    

                                               
{'eval_loss': nan, 'eval_runtime': 135.8684, 'eval_samples_per_second': 0.765, 'eval_steps_per_second': 0.096, 'eval_num_tokens': 320410.0, 'eval_completions/mean_length': 157.42307692307693, 'eval_completions/min_length': 39.53846153846154, 'eval_completions/max_length': 349.84615384615387, 'eval_completions/clipped_ratio': 0.009615384615384616, 'eval_completions/mean_terminated_length': 153.90934166541467, 'eval_completions/min_terminated_length': 39.53846153846154, 'eval_completions/max_terminated_length': 332.6923076923077, 'eval_rewards/meter/mean': 0.4222999421449808, 'eval_rewards/meter/std': 0.3532765622322376, 'eval_rewards/count_adherence/mean': 0.9509367025815524, 'eval_rewards/count_adherence/std': 0.06137575008548223, 'eval_rewards/arabic_clean/mean': 0.8557692307692307, 'eval_rewards/arabic_clean/std': 0.29103099841337937, 'eval_rewards/lexical_plausibility/mean': 0.9822093019118676, 'eval_rewards/lexical_plausibility/std': 0.037934203882916614, 'eval_rewards/judge_quality/mean': 0.3713461515995172, 'eval_rewards/judge_quality/std': 0.1981737602215547, 'eval_rewards/repeat_penalty/mean': 0.9721828103065491, 'eval_rewards/repeat_penalty/std': 0.0459989532828331, 'eval_rewards/near_duplicate_penalty/mean': 0.9884350712482746, 'eval_rewards/near_duplicate_penalty/std': 0.015634029432056613, 'eval_rewards/opening_diversity/mean': 0.9927884615384616, 'eval_rewards/opening_diversity/std': 0.020397310646680687, 'eval_rewards/distinct_2/mean': 0.9909096543605511, 'eval_rewards/distinct_2/std': 0.020599648750458773, 'eval_rewards/total_composite/mean': 0.13668288147220245, 'eval_rewards/total_composite/std': 0.16516263782978058, 'eval_reward': 0.13668288147220245, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.16735697021851173, 'eval_sampling/sampling_logp_difference/max': 1.3967525775615985, 'eval_sampling/importance_sampling_ratio/min': 0.2492638390797835, 'eval_sampling/importance_sampling_ratio/mean': 1.0539659078304584, 'eval_sampling/importance_sampling_ratio/max': 1.7433728346457849, 'eval_entropy': 2.9130990413519053, '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': 0.13668288147220245, 'eval_reward_meter_mean': 0.4222999421449808, 'eval_reward_meter_std': 0.3532765622322376, 'eval_reward_count_adherence_mean': 0.9509367025815524, 'eval_reward_count_adherence_std': 0.06137575008548223, 'eval_reward_arabic_clean_mean': 0.8557692307692307, 'eval_reward_arabic_clean_std': 0.29103099841337937, 'eval_reward_lexical_plausibility_mean': 0.9822093019118676, 'eval_reward_lexical_plausibility_std': 0.037934203882916614, 'eval_reward_repeat_penalty_mean': 0.9721828103065491, 'eval_reward_repeat_penalty_std': 0.0459989532828331, 'eval_reward_near_duplicate_penalty_mean': 0.9884350712482746, 'eval_reward_near_duplicate_penalty_std': 0.015634029432056613, 'eval_reward_opening_diversity_mean': 0.9927884615384616, 'eval_reward_opening_diversity_std': 0.020397310646680687, 'eval_reward_distinct_2_mean': 0.9909096543605511, 'eval_reward_distinct_2_std': 0.020599648750458773, 'eval_reward_judge_quality_mean': 0.3713461515995172, 'eval_reward_judge_quality_std': 0.1981737602215547, 'eval_reward_total_composite_mean': 0.13668288147220245, 'eval_reward_total_composite_std': 0.16516263782978058, 'epoch': 0.01}

  5%|▍         | 150/3300 [33:14<8:38:17,  9.87s/it]

100%|██████████| 13/13 [02:05<00:00, 10.79s/it]

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

  5%|▍         | 151/3300 [33:27<45:07:10, 51.58s/it]2026-04-12 11:19:39,448 | INFO | train_grpo_train | metrics_logged mode=train step=151

                                                     
{'loss': -0.0141, 'grad_norm': 14.000923156738281, 'learning_rate': 9.545454545454547e-06, 'num_tokens': 321978.0, 'completions/mean_length': 45.0, 'completions/min_length': 31.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.0, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.6058230996131897, 'rewards/meter/std': 0.43810972571372986, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6799999475479126, 'rewards/judge_quality/std': 0.2747986316680908, 'rewards/repeat_penalty/mean': 0.9692283868789673, 'rewards/repeat_penalty/std': 0.03055117465555668, 'rewards/near_duplicate_penalty/mean': 0.9692283868789673, 'rewards/near_duplicate_penalty/std': 0.03055117465555668, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.45389193296432495, 'rewards/total_composite/std': 0.40621599555015564, 'reward': 0.45389193296432495, 'reward_std': 0.40621596574783325, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.200367271900177, 'sampling/sampling_logp_difference/max': 1.2588658332824707, 'sampling/importance_sampling_ratio/min': 0.28397589921951294, 'sampling/importance_sampling_ratio/mean': 1.0334672927856445, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9045850485563278, 'clip_ratio/low_mean': 0.13465113565325737, 'clip_ratio/low_min': 0.13465113565325737, 'clip_ratio/high_mean': 0.0493504973128438, 'clip_ratio/high_max': 0.0493504973128438, 'clip_ratio/region_mean': 0.18400163296610117, 'reward_total_mean': 0.45389193296432495, 'reward_meter_mean': 0.6058230996131897, 'reward_meter_std': 0.43810972571372986, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9692283868789673, 'reward_repeat_penalty_std': 0.03055117465555668, 'reward_near_duplicate_penalty_mean': 0.9692283868789673, 'reward_near_duplicate_penalty_std': 0.03055117465555668, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6799999475479126, 'reward_judge_quality_std': 0.2747986316680908, 'reward_total_composite_mean': 0.45389193296432495, 'reward_total_composite_std': 0.40621599555015564, 'epoch': 0.01}

  5%|▍         | 151/3300 [33:27<45:07:10, 51.58s/it]INFO 04-12 11:19:39 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 152/3300 [33:37<34:15:02, 39.17s/it]2026-04-12 11:19:49,653 | INFO | train_grpo_train | metrics_logged mode=train step=152

                                                     
{'loss': 0.0195, 'grad_norm': 8.529622077941895, 'learning_rate': 9.542424242424242e-06, 'num_tokens': 324395.0, 'completions/mean_length': 111.125, 'completions/min_length': 99.0, 'completions/max_length': 129.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 111.125, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 129.0, 'rewards/meter/mean': 0.8794379234313965, 'rewards/meter/std': 0.19269824028015137, '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/lexical_plausibility/mean': 0.9983552694320679, 'rewards/lexical_plausibility/std': 0.004652021918445826, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.9949206709861755, 'rewards/repeat_penalty/std': 0.0037612332962453365, 'rewards/near_duplicate_penalty/mean': 0.9949206709861755, 'rewards/near_duplicate_penalty/std': 0.0037612332962453365, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.20532241463661194, 'rewards/total_composite/std': 0.1613740772008896, 'reward': 0.20532241463661194, 'reward_std': 0.16137409210205078, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21348437666893005, 'sampling/sampling_logp_difference/max': 1.2490053176879883, 'sampling/importance_sampling_ratio/min': 0.2867899239063263, 'sampling/importance_sampling_ratio/mean': 1.0697646141052246, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.0969585478305817, 'clip_ratio/low_mean': 0.07440471276640892, 'clip_ratio/low_min': 0.07440471276640892, 'clip_ratio/high_mean': 0.08561559952795506, 'clip_ratio/high_max': 0.08561559952795506, 'clip_ratio/region_mean': 0.16002031229436398, 'reward_total_mean': 0.20532241463661194, 'reward_meter_mean': 0.8794379234313965, 'reward_meter_std': 0.19269824028015137, '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_lexical_plausibility_mean': 0.9983552694320679, 'reward_lexical_plausibility_std': 0.004652021918445826, 'reward_repeat_penalty_mean': 0.9949206709861755, 'reward_repeat_penalty_std': 0.0037612332962453365, 'reward_near_duplicate_penalty_mean': 0.9949206709861755, 'reward_near_duplicate_penalty_std': 0.0037612332962453365, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.20532241463661194, 'reward_total_composite_std': 0.1613740772008896, 'epoch': 0.01}

  5%|▍         | 152/3300 [33:37<34:15:02, 39.17s/it]INFO 04-12 11:19:49 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 153/3300 [33:48<26:55:54, 30.81s/it]2026-04-12 11:20:00,955 | INFO | train_grpo_train | metrics_logged mode=train step=153

                                                     
{'loss': 0.0369, 'grad_norm': 9.671324729919434, 'learning_rate': 9.539393939393941e-06, 'num_tokens': 326571.0, 'completions/mean_length': 93.0, 'completions/min_length': 68.0, 'completions/max_length': 136.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 93.0, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 136.0, 'rewards/meter/mean': 0.5104454159736633, 'rewards/meter/std': 0.2965039014816284, '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/lexical_plausibility/mean': 0.9899665713310242, 'rewards/lexical_plausibility/std': 0.028378864750266075, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.9915590882301331, 'rewards/repeat_penalty/std': 0.015696531161665916, 'rewards/near_duplicate_penalty/mean': 0.9962067008018494, 'rewards/near_duplicate_penalty/std': 0.004112233407795429, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.995309591293335, 'rewards/distinct_2/std': 0.013266542926430702, 'rewards/total_composite/mean': 0.11554411798715591, 'rewards/total_composite/std': 0.11518596112728119, 'reward': 0.11554411798715591, 'reward_std': 0.11518596857786179, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23454831540584564, 'sampling/sampling_logp_difference/max': 2.2007312774658203, 'sampling/importance_sampling_ratio/min': 0.11072216182947159, 'sampling/importance_sampling_ratio/mean': 1.035933256149292, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.980690121650696, 'clip_ratio/low_mean': 0.13396853860467672, 'clip_ratio/low_min': 0.13396853860467672, 'clip_ratio/high_mean': 0.05198572017252445, 'clip_ratio/high_max': 0.05198572017252445, 'clip_ratio/region_mean': 0.18595425877720118, 'reward_total_mean': 0.11554411798715591, 'reward_meter_mean': 0.5104454159736633, 'reward_meter_std': 0.2965039014816284, '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_lexical_plausibility_mean': 0.9899665713310242, 'reward_lexical_plausibility_std': 0.028378864750266075, 'reward_repeat_penalty_mean': 0.9915590882301331, 'reward_repeat_penalty_std': 0.015696531161665916, 'reward_near_duplicate_penalty_mean': 0.9962067008018494, 'reward_near_duplicate_penalty_std': 0.004112233407795429, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.995309591293335, 'reward_distinct_2_std': 0.013266542926430702, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.11554411798715591, 'reward_total_composite_std': 0.11518596112728119, 'epoch': 0.01}

  5%|▍         | 153/3300 [33:48<26:55:54, 30.81s/it]INFO 04-12 11:20:01 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 154/3300 [33:58<21:28:27, 24.57s/it]2026-04-12 11:20:10,985 | INFO | train_grpo_train | metrics_logged mode=train step=154

                                                     
{'loss': 0.0103, 'grad_norm': 20.035627365112305, 'learning_rate': 9.536363636363637e-06, 'num_tokens': 328056.0, 'completions/mean_length': 25.625, 'completions/min_length': 17.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.625, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.7828391790390015, 'rewards/meter/std': 0.30151939392089844, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.8999863862991333, 'rewards/lexical_plausibility/std': 0.11114007234573364, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.26569312810897827, 'rewards/total_composite/std': 0.15375202894210815, 'reward': 0.26569312810897827, 'reward_std': 0.15375201404094696, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2073061317205429, 'sampling/sampling_logp_difference/max': 1.3186922073364258, 'sampling/importance_sampling_ratio/min': 0.2674849033355713, 'sampling/importance_sampling_ratio/mean': 1.0439095497131348, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.5544720590114594, 'clip_ratio/low_mean': 0.060484389774501324, 'clip_ratio/low_min': 0.060484389774501324, 'clip_ratio/high_mean': 0.13336967583745718, 'clip_ratio/high_max': 0.13336967583745718, 'clip_ratio/region_mean': 0.1938540656119585, 'reward_total_mean': 0.26569312810897827, 'reward_meter_mean': 0.7828391790390015, 'reward_meter_std': 0.30151939392089844, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.8999863862991333, 'reward_lexical_plausibility_std': 0.11114007234573364, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.26569312810897827, 'reward_total_composite_std': 0.15375202894210815, 'epoch': 0.01}

  5%|▍         | 154/3300 [33:58<21:28:27, 24.57s/it]INFO 04-12 11:20:11 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 155/3300 [34:07<17:22:53, 19.90s/it]2026-04-12 11:20:19,966 | INFO | train_grpo_train | metrics_logged mode=train step=155

                                                     
{'loss': 0.158, 'grad_norm': 17.03345489501953, 'learning_rate': 9.533333333333334e-06, 'num_tokens': 329831.0, 'completions/mean_length': 38.875, 'completions/min_length': 18.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.875, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.32180482149124146, 'rewards/meter/std': 0.37046319246292114, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.6112217903137207, 'rewards/lexical_plausibility/std': 0.2747129499912262, 'rewards/judge_quality/mean': 0.5512499809265137, 'rewards/judge_quality/std': 0.3296508193016052, 'rewards/repeat_penalty/mean': 0.8932257890701294, 'rewards/repeat_penalty/std': 0.15036876499652863, 'rewards/near_duplicate_penalty/mean': 0.9934245347976685, 'rewards/near_duplicate_penalty/std': 0.01804819144308567, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9886877536773682, 'rewards/distinct_2/std': 0.031995780766010284, 'rewards/total_composite/mean': 0.13939277827739716, 'rewards/total_composite/std': 0.16900645196437836, 'reward': 0.13939277827739716, 'reward_std': 0.16900645196437836, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.228168785572052, 'sampling/sampling_logp_difference/max': 1.8393783569335938, 'sampling/importance_sampling_ratio/min': 0.158916175365448, 'sampling/importance_sampling_ratio/mean': 1.0233756303787231, 'sampling/importance_sampling_ratio/max': 1.908592700958252, 'entropy': 2.470901608467102, 'clip_ratio/low_mean': 0.11315600201487541, 'clip_ratio/low_min': 0.11315600201487541, 'clip_ratio/high_mean': 0.09409722313284874, 'clip_ratio/high_max': 0.09409722313284874, 'clip_ratio/region_mean': 0.20725322514772415, 'reward_total_mean': 0.13939277827739716, 'reward_meter_mean': 0.32180482149124146, 'reward_meter_std': 0.37046319246292114, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.6112217903137207, 'reward_lexical_plausibility_std': 0.2747129499912262, 'reward_repeat_penalty_mean': 0.8932257890701294, 'reward_repeat_penalty_std': 0.15036876499652863, 'reward_near_duplicate_penalty_mean': 0.9934245347976685, 'reward_near_duplicate_penalty_std': 0.01804819144308567, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9886877536773682, 'reward_distinct_2_std': 0.031995780766010284, 'reward_judge_quality_mean': 0.5512499809265137, 'reward_judge_quality_std': 0.3296508193016052, 'reward_total_composite_mean': 0.13939277827739716, 'reward_total_composite_std': 0.16900645196437836, 'epoch': 0.01}

  5%|▍         | 155/3300 [34:07<17:22:53, 19.90s/it]INFO 04-12 11:20:20 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 156/3300 [34:18<14:56:18, 17.10s/it]2026-04-12 11:20:30,553 | INFO | train_grpo_train | metrics_logged mode=train step=156

                                                     
{'loss': -0.0176, 'grad_norm': 11.074264526367188, 'learning_rate': 9.530303030303031e-06, 'num_tokens': 331924.0, 'completions/mean_length': 78.625, 'completions/min_length': 70.0, 'completions/max_length': 94.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 78.625, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.47446614503860474, 'rewards/meter/std': 0.16489452123641968, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.9808568358421326, 'rewards/repeat_penalty/std': 0.017523130401968956, 'rewards/near_duplicate_penalty/mean': 0.9856008887290955, 'rewards/near_duplicate_penalty/std': 0.011811843141913414, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.995192289352417, 'rewards/distinct_2/std': 0.013598216697573662, 'rewards/total_composite/mean': 0.1452259123325348, 'rewards/total_composite/std': 0.0769054964184761, 'reward': 0.1452259123325348, 'reward_std': 0.07690548896789551, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22640937566757202, 'sampling/sampling_logp_difference/max': 1.5695104598999023, 'sampling/importance_sampling_ratio/min': 0.2081470489501953, 'sampling/importance_sampling_ratio/mean': 1.0527526140213013, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.949790358543396, 'clip_ratio/low_mean': 0.07967651262879372, 'clip_ratio/low_min': 0.07967651262879372, 'clip_ratio/high_mean': 0.12765267305076122, 'clip_ratio/high_max': 0.12765267305076122, 'clip_ratio/region_mean': 0.20732918567955494, 'reward_total_mean': 0.1452259123325348, 'reward_meter_mean': 0.47446614503860474, 'reward_meter_std': 0.16489452123641968, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9808568358421326, 'reward_repeat_penalty_std': 0.017523130401968956, 'reward_near_duplicate_penalty_mean': 0.9856008887290955, 'reward_near_duplicate_penalty_std': 0.011811843141913414, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.995192289352417, 'reward_distinct_2_std': 0.013598216697573662, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.1452259123325348, 'reward_total_composite_std': 0.0769054964184761, 'epoch': 0.01}

  5%|▍         | 156/3300 [34:18<14:56:18, 17.10s/it]INFO 04-12 11:20:30 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 157/3300 [34:27<12:54:26, 14.78s/it]2026-04-12 11:20:39,923 | INFO | train_grpo_train | metrics_logged mode=train step=157

                                                     
{'loss': 0.1223, 'grad_norm': 24.670238494873047, 'learning_rate': 9.527272727272729e-06, 'num_tokens': 333224.0, 'completions/mean_length': 20.5, 'completions/min_length': 17.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.5, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.8423395156860352, 'rewards/meter/std': 0.34059402346611023, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8350000381469727, 'rewards/judge_quality/std': 0.27717968821525574, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.690671980381012, 'rewards/total_composite/std': 0.38380205631256104, 'reward': 0.690671980381012, 'reward_std': 0.38380205631256104, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1779317408800125, 'sampling/sampling_logp_difference/max': 1.5056533813476562, 'sampling/importance_sampling_ratio/min': 0.22187228500843048, 'sampling/importance_sampling_ratio/mean': 1.0350496768951416, 'sampling/importance_sampling_ratio/max': 1.829878568649292, 'entropy': 1.6537972539663315, 'clip_ratio/low_mean': 0.033409091643989086, 'clip_ratio/low_min': 0.033409091643989086, 'clip_ratio/high_mean': 0.08769813040271401, 'clip_ratio/high_max': 0.08769813040271401, 'clip_ratio/region_mean': 0.1211072220467031, 'reward_total_mean': 0.690671980381012, 'reward_meter_mean': 0.8423395156860352, 'reward_meter_std': 0.34059402346611023, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8350000381469727, 'reward_judge_quality_std': 0.27717968821525574, 'reward_total_composite_mean': 0.690671980381012, 'reward_total_composite_std': 0.38380205631256104, 'epoch': 0.01}

  5%|▍         | 157/3300 [34:27<12:54:26, 14.78s/it]INFO 04-12 11:20:40 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 158/3300 [34:37<11:40:38, 13.38s/it]2026-04-12 11:20:50,061 | INFO | train_grpo_train | metrics_logged mode=train step=158

                                                     
{'loss': 0.0628, 'grad_norm': 12.094988822937012, 'learning_rate': 9.524242424242424e-06, 'num_tokens': 335028.0, 'completions/mean_length': 65.5, 'completions/min_length': 30.0, 'completions/max_length': 104.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.5, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 104.0, 'rewards/meter/mean': 0.5528392195701599, 'rewards/meter/std': 0.39363592863082886, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.2357022762298584, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9114699363708496, 'rewards/lexical_plausibility/std': 0.16236990690231323, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.10690449178218842, 'rewards/repeat_penalty/mean': 0.9983820915222168, 'rewards/repeat_penalty/std': 0.0038890535943210125, 'rewards/near_duplicate_penalty/mean': 0.9983820915222168, 'rewards/near_duplicate_penalty/std': 0.0038890535943210125, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.11301341652870178, 'rewards/total_composite/std': 0.10902532935142517, 'reward': 0.11301341652870178, 'reward_std': 0.10902532935142517, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23686851561069489, 'sampling/sampling_logp_difference/max': 1.5974016189575195, 'sampling/importance_sampling_ratio/min': 0.20242181420326233, 'sampling/importance_sampling_ratio/mean': 1.0388325452804565, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.01512011885643, 'clip_ratio/low_mean': 0.1251815352588892, 'clip_ratio/low_min': 0.1251815352588892, 'clip_ratio/high_mean': 0.08592366427183151, 'clip_ratio/high_max': 0.08592366427183151, 'clip_ratio/region_mean': 0.2111051995307207, 'reward_total_mean': 0.11301341652870178, 'reward_meter_mean': 0.5528392195701599, 'reward_meter_std': 0.39363592863082886, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.2357022762298584, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9114699363708496, 'reward_lexical_plausibility_std': 0.16236990690231323, 'reward_repeat_penalty_mean': 0.9983820915222168, 'reward_repeat_penalty_std': 0.0038890535943210125, 'reward_near_duplicate_penalty_mean': 0.9983820915222168, 'reward_near_duplicate_penalty_std': 0.0038890535943210125, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.10690449178218842, 'reward_total_composite_mean': 0.11301341652870178, 'reward_total_composite_std': 0.10902532935142517, 'epoch': 0.01}

  5%|▍         | 158/3300 [34:38<11:40:38, 13.38s/it]INFO 04-12 11:20:50 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 159/3300 [34:47<10:40:15, 12.23s/it]2026-04-12 11:20:59,574 | INFO | train_grpo_train | metrics_logged mode=train step=159

                                                     
{'loss': 0.0045, 'grad_norm': 13.970623970031738, 'learning_rate': 9.521212121212121e-06, 'num_tokens': 337020.0, 'completions/mean_length': 44.0, 'completions/min_length': 33.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.0, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.7165439128875732, 'rewards/meter/std': 0.3721120059490204, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4587499797344208, 'rewards/judge_quality/std': 0.192831352353096, 'rewards/repeat_penalty/mean': 0.9932456016540527, 'rewards/repeat_penalty/std': 0.009441905654966831, 'rewards/near_duplicate_penalty/mean': 0.9932456016540527, 'rewards/near_duplicate_penalty/std': 0.009441905654966831, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3526611328125, 'rewards/total_composite/std': 0.267297625541687, 'reward': 0.3526611328125, 'reward_std': 0.267297625541687, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17874982953071594, 'sampling/sampling_logp_difference/max': 1.375375747680664, 'sampling/importance_sampling_ratio/min': 0.2527446150779724, 'sampling/importance_sampling_ratio/mean': 1.0489211082458496, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8119719177484512, 'clip_ratio/low_mean': 0.10304794274270535, 'clip_ratio/low_min': 0.10304794274270535, 'clip_ratio/high_mean': 0.06766779348254204, 'clip_ratio/high_max': 0.06766779348254204, 'clip_ratio/region_mean': 0.17071573622524738, 'reward_total_mean': 0.3526611328125, 'reward_meter_mean': 0.7165439128875732, 'reward_meter_std': 0.3721120059490204, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9932456016540527, 'reward_repeat_penalty_std': 0.009441905654966831, 'reward_near_duplicate_penalty_mean': 0.9932456016540527, 'reward_near_duplicate_penalty_std': 0.009441905654966831, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4587499797344208, 'reward_judge_quality_std': 0.192831352353096, 'reward_total_composite_mean': 0.3526611328125, 'reward_total_composite_std': 0.267297625541687, 'epoch': 0.01}

  5%|▍         | 159/3300 [34:47<10:40:15, 12.23s/it]INFO 04-12 11:20:59 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 160/3300 [34:59<10:29:38, 12.03s/it]2026-04-12 11:21:11,141 | INFO | train_grpo_train | metrics_logged mode=train step=160

                                                     
{'loss': -0.0205, 'grad_norm': 15.368523597717285, 'learning_rate': 9.518181818181819e-06, 'num_tokens': 338994.0, 'completions/mean_length': 65.75, 'completions/min_length': 47.0, 'completions/max_length': 82.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.75, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.3450881838798523, 'rewards/meter/std': 0.24610362946987152, '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/lexical_plausibility/mean': 0.9259542226791382, 'rewards/lexical_plausibility/std': 0.1133645623922348, 'rewards/judge_quality/mean': 0.3712499737739563, 'rewards/judge_quality/std': 0.23943014442920685, 'rewards/repeat_penalty/mean': 0.9626677632331848, 'rewards/repeat_penalty/std': 0.04355334863066673, 'rewards/near_duplicate_penalty/mean': 0.9872620105743408, 'rewards/near_duplicate_penalty/std': 0.011765980161726475, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9748167991638184, 'rewards/distinct_2/std': 0.03487076237797737, 'rewards/total_composite/mean': 0.09414747357368469, 'rewards/total_composite/std': 0.05192834511399269, 'reward': 0.09414747357368469, 'reward_std': 0.05192834511399269, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19198961555957794, 'sampling/sampling_logp_difference/max': 1.6694979667663574, 'sampling/importance_sampling_ratio/min': 0.18834158778190613, 'sampling/importance_sampling_ratio/mean': 1.0056805610656738, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3271846771240234, 'clip_ratio/low_mean': 0.1068890830501914, 'clip_ratio/low_min': 0.1068890830501914, 'clip_ratio/high_mean': 0.07343112863600254, 'clip_ratio/high_max': 0.07343112863600254, 'clip_ratio/region_mean': 0.18032021168619394, 'reward_total_mean': 0.09414747357368469, 'reward_meter_mean': 0.3450881838798523, 'reward_meter_std': 0.24610362946987152, '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_lexical_plausibility_mean': 0.9259542226791382, 'reward_lexical_plausibility_std': 0.1133645623922348, 'reward_repeat_penalty_mean': 0.9626677632331848, 'reward_repeat_penalty_std': 0.04355334863066673, 'reward_near_duplicate_penalty_mean': 0.9872620105743408, 'reward_near_duplicate_penalty_std': 0.011765980161726475, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9748167991638184, 'reward_distinct_2_std': 0.03487076237797737, 'reward_judge_quality_mean': 0.3712499737739563, 'reward_judge_quality_std': 0.23943014442920685, 'reward_total_composite_mean': 0.09414747357368469, 'reward_total_composite_std': 0.05192834511399269, 'epoch': 0.01}

  5%|▍         | 160/3300 [34:59<10:29:38, 12.03s/it]INFO 04-12 11:21:11 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 161/3300 [35:09<10:02:51, 11.52s/it]2026-04-12 11:21:21,489 | INFO | train_grpo_train | metrics_logged mode=train step=161

                                                     
{'loss': 0.0945, 'grad_norm': 12.157179832458496, 'learning_rate': 9.515151515151516e-06, 'num_tokens': 340967.0, 'completions/mean_length': 68.625, 'completions/min_length': 57.0, 'completions/max_length': 91.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.625, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 91.0, 'rewards/meter/mean': 0.8782539367675781, 'rewards/meter/std': 0.16990414261817932, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9693520069122314, 'rewards/lexical_plausibility/std': 0.07608518749475479, 'rewards/judge_quality/mean': 0.36249998211860657, 'rewards/judge_quality/std': 0.1642080545425415, 'rewards/repeat_penalty/mean': 0.9892227649688721, 'rewards/repeat_penalty/std': 0.009906812570989132, 'rewards/near_duplicate_penalty/mean': 0.9892227649688721, 'rewards/near_duplicate_penalty/std': 0.009906812570989132, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.32068121433258057, 'rewards/total_composite/std': 0.14033915102481842, 'reward': 0.32068121433258057, 'reward_std': 0.14033915102481842, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22978489100933075, 'sampling/sampling_logp_difference/max': 1.396162509918213, 'sampling/importance_sampling_ratio/min': 0.24754510819911957, 'sampling/importance_sampling_ratio/mean': 1.043877124786377, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.8486823439598083, 'clip_ratio/low_mean': 0.05545760318636894, 'clip_ratio/low_min': 0.05545760318636894, 'clip_ratio/high_mean': 0.15594460628926754, 'clip_ratio/high_max': 0.15594460628926754, 'clip_ratio/region_mean': 0.21140220947563648, 'reward_total_mean': 0.32068121433258057, 'reward_meter_mean': 0.8782539367675781, 'reward_meter_std': 0.16990414261817932, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9693520069122314, 'reward_lexical_plausibility_std': 0.07608518749475479, 'reward_repeat_penalty_mean': 0.9892227649688721, 'reward_repeat_penalty_std': 0.009906812570989132, 'reward_near_duplicate_penalty_mean': 0.9892227649688721, 'reward_near_duplicate_penalty_std': 0.009906812570989132, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.36249998211860657, 'reward_judge_quality_std': 0.1642080545425415, 'reward_total_composite_mean': 0.32068121433258057, 'reward_total_composite_std': 0.14033915102481842, 'epoch': 0.01}

  5%|▍         | 161/3300 [35:09<10:02:51, 11.52s/it]INFO 04-12 11:21:21 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 162/3300 [35:20<9:59:35, 11.46s/it] 2026-04-12 11:21:32,809 | INFO | train_grpo_train | metrics_logged mode=train step=162

                                                    
{'loss': 0.1007, 'grad_norm': 7.251101016998291, 'learning_rate': 9.512121212121213e-06, 'num_tokens': 344016.0, 'completions/mean_length': 173.125, 'completions/min_length': 147.0, 'completions/max_length': 200.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 173.125, 'completions/min_terminated_length': 147.0, 'completions/max_terminated_length': 200.0, 'rewards/meter/mean': 0.6534852981567383, 'rewards/meter/std': 0.37754520773887634, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.05050762742757797, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9472790956497192, 'rewards/lexical_plausibility/std': 0.0891343355178833, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.9930524826049805, 'rewards/repeat_penalty/std': 0.0058747329749166965, 'rewards/near_duplicate_penalty/mean': 0.995485782623291, 'rewards/near_duplicate_penalty/std': 0.0031709817703813314, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.997565746307373, 'rewards/distinct_2/std': 0.006885172799229622, 'rewards/total_composite/mean': 0.18729378283023834, 'rewards/total_composite/std': 0.1243356391787529, 'reward': 0.18729378283023834, 'reward_std': 0.1243356391787529, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23811013996601105, 'sampling/sampling_logp_difference/max': 1.5276861190795898, 'sampling/importance_sampling_ratio/min': 0.21703729033470154, 'sampling/importance_sampling_ratio/mean': 1.0599641799926758, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.803977072238922, 'clip_ratio/low_mean': 0.06561482138931751, 'clip_ratio/low_min': 0.06561482138931751, 'clip_ratio/high_mean': 0.14769021049141884, 'clip_ratio/high_max': 0.14769021049141884, 'clip_ratio/region_mean': 0.21330503188073635, 'reward_total_mean': 0.18729378283023834, 'reward_meter_mean': 0.6534852981567383, 'reward_meter_std': 0.37754520773887634, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.05050762742757797, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9472790956497192, 'reward_lexical_plausibility_std': 0.0891343355178833, 'reward_repeat_penalty_mean': 0.9930524826049805, 'reward_repeat_penalty_std': 0.0058747329749166965, 'reward_near_duplicate_penalty_mean': 0.995485782623291, 'reward_near_duplicate_penalty_std': 0.0031709817703813314, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.997565746307373, 'reward_distinct_2_std': 0.006885172799229622, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.18729378283023834, 'reward_total_composite_std': 0.1243356391787529, 'epoch': 0.01}

  5%|▍         | 162/3300 [35:20<9:59:35, 11.46s/it]INFO 04-12 11:21:33 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 163/3300 [35:30<9:38:22, 11.06s/it]2026-04-12 11:21:42,930 | INFO | train_grpo_train | metrics_logged mode=train step=163

                                                    
{'loss': 0.0867, 'grad_norm': 15.229354858398438, 'learning_rate': 9.50909090909091e-06, 'num_tokens': 345639.0, 'completions/mean_length': 42.875, 'completions/min_length': 37.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.875, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.605192244052887, 'rewards/meter/std': 0.34179389476776123, '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/lexical_plausibility/mean': 0.9881524443626404, 'rewards/lexical_plausibility/std': 0.01754402369260788, 'rewards/judge_quality/mean': 0.3375000059604645, 'rewards/judge_quality/std': 0.155264750123024, 'rewards/repeat_penalty/mean': 0.9992201328277588, 'rewards/repeat_penalty/std': 0.0022057765163481236, 'rewards/near_duplicate_penalty/mean': 0.9992201328277588, 'rewards/near_duplicate_penalty/std': 0.0022057765163481236, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2018437683582306, 'rewards/total_composite/std': 0.16896285116672516, 'reward': 0.2018437683582306, 'reward_std': 0.16896285116672516, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23416592180728912, 'sampling/sampling_logp_difference/max': 1.3484859466552734, 'sampling/importance_sampling_ratio/min': 0.25963306427001953, 'sampling/importance_sampling_ratio/mean': 1.040067195892334, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.308599442243576, 'clip_ratio/low_mean': 0.10402058809995651, 'clip_ratio/low_min': 0.10402058809995651, 'clip_ratio/high_mean': 0.1089235981926322, 'clip_ratio/high_max': 0.1089235981926322, 'clip_ratio/region_mean': 0.2129441862925887, 'reward_total_mean': 0.2018437683582306, 'reward_meter_mean': 0.605192244052887, 'reward_meter_std': 0.34179389476776123, '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_lexical_plausibility_mean': 0.9881524443626404, 'reward_lexical_plausibility_std': 0.01754402369260788, 'reward_repeat_penalty_mean': 0.9992201328277588, 'reward_repeat_penalty_std': 0.0022057765163481236, 'reward_near_duplicate_penalty_mean': 0.9992201328277588, 'reward_near_duplicate_penalty_std': 0.0022057765163481236, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3375000059604645, 'reward_judge_quality_std': 0.155264750123024, 'reward_total_composite_mean': 0.2018437683582306, 'reward_total_composite_std': 0.16896285116672516, 'epoch': 0.01}

  5%|▍         | 163/3300 [35:30<9:38:22, 11.06s/it]INFO 04-12 11:21:43 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 164/3300 [35:40<9:11:47, 10.56s/it]2026-04-12 11:21:52,317 | INFO | train_grpo_train | metrics_logged mode=train step=164

                                                    
{'loss': 0.0515, 'grad_norm': 16.58517837524414, 'learning_rate': 9.506060606060606e-06, 'num_tokens': 347431.0, 'completions/mean_length': 47.0, 'completions/min_length': 34.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.0, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.3476116955280304, 'rewards/meter/std': 0.31187713146209717, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9825247526168823, 'rewards/lexical_plausibility/std': 0.028884755447506905, 'rewards/judge_quality/mean': 0.5549999475479126, 'rewards/judge_quality/std': 0.22790977358818054, 'rewards/repeat_penalty/mean': 0.9899319410324097, 'rewards/repeat_penalty/std': 0.015661869198083878, 'rewards/near_duplicate_penalty/mean': 0.9899319410324097, 'rewards/near_duplicate_penalty/std': 0.015661869198083878, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.21823325753211975, 'rewards/total_composite/std': 0.23379917442798615, 'reward': 0.21823325753211975, 'reward_std': 0.23379915952682495, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23287346959114075, 'sampling/sampling_logp_difference/max': 1.274601936340332, 'sampling/importance_sampling_ratio/min': 0.2795422077178955, 'sampling/importance_sampling_ratio/mean': 1.0410550832748413, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.10340116918087, 'clip_ratio/low_mean': 0.12830465659499168, 'clip_ratio/low_min': 0.12830465659499168, 'clip_ratio/high_mean': 0.07553191483020782, 'clip_ratio/high_max': 0.07553191483020782, 'clip_ratio/region_mean': 0.2038365714251995, 'reward_total_mean': 0.21823325753211975, 'reward_meter_mean': 0.3476116955280304, 'reward_meter_std': 0.31187713146209717, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9825247526168823, 'reward_lexical_plausibility_std': 0.028884755447506905, 'reward_repeat_penalty_mean': 0.9899319410324097, 'reward_repeat_penalty_std': 0.015661869198083878, 'reward_near_duplicate_penalty_mean': 0.9899319410324097, 'reward_near_duplicate_penalty_std': 0.015661869198083878, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5549999475479126, 'reward_judge_quality_std': 0.22790977358818054, 'reward_total_composite_mean': 0.21823325753211975, 'reward_total_composite_std': 0.23379917442798615, 'epoch': 0.01}

  5%|▍         | 164/3300 [35:40<9:11:47, 10.56s/it]INFO 04-12 11:21:52 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 165/3300 [35:51<9:29:04, 10.89s/it]2026-04-12 11:22:03,983 | INFO | train_grpo_train | metrics_logged mode=train step=165

                                                    
{'loss': 0.1233, 'grad_norm': 8.282011032104492, 'learning_rate': 9.503030303030303e-06, 'num_tokens': 349894.0, 'completions/mean_length': 130.875, 'completions/min_length': 110.0, 'completions/max_length': 231.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 130.875, 'completions/min_terminated_length': 110.0, 'completions/max_terminated_length': 231.0, 'rewards/meter/mean': 0.7863786816596985, 'rewards/meter/std': 0.2986995279788971, 'rewards/count_adherence/mean': 0.9791666269302368, 'rewards/count_adherence/std': 0.0589255727827549, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9920691251754761, 'rewards/lexical_plausibility/std': 0.018583906814455986, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.9819804430007935, 'rewards/repeat_penalty/std': 0.02719215303659439, 'rewards/near_duplicate_penalty/mean': 0.9935469627380371, 'rewards/near_duplicate_penalty/std': 0.006173244211822748, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9882498979568481, 'rewards/distinct_2/std': 0.02176465094089508, 'rewards/total_composite/mean': 0.22972571849822998, 'rewards/total_composite/std': 0.15144440531730652, 'reward': 0.22972571849822998, 'reward_std': 0.15144439041614532, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23464597761631012, 'sampling/sampling_logp_difference/max': 1.9085321426391602, 'sampling/importance_sampling_ratio/min': 0.14829790592193604, 'sampling/importance_sampling_ratio/mean': 1.055251955986023, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.7573797404766083, 'clip_ratio/low_mean': 0.08293617330491543, 'clip_ratio/low_min': 0.08293617330491543, 'clip_ratio/high_mean': 0.09958924725651741, 'clip_ratio/high_max': 0.09958924725651741, 'clip_ratio/region_mean': 0.18252542056143284, 'reward_total_mean': 0.22972571849822998, 'reward_meter_mean': 0.7863786816596985, 'reward_meter_std': 0.2986995279788971, 'reward_count_adherence_mean': 0.9791666269302368, 'reward_count_adherence_std': 0.0589255727827549, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9920691251754761, 'reward_lexical_plausibility_std': 0.018583906814455986, 'reward_repeat_penalty_mean': 0.9819804430007935, 'reward_repeat_penalty_std': 0.02719215303659439, 'reward_near_duplicate_penalty_mean': 0.9935469627380371, 'reward_near_duplicate_penalty_std': 0.006173244211822748, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9882498979568481, 'reward_distinct_2_std': 0.02176465094089508, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.22972571849822998, 'reward_total_composite_std': 0.15144440531730652, 'epoch': 0.01}

  5%|▌         | 165/3300 [35:51<9:29:04, 10.89s/it]INFO 04-12 11:22:04 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 166/3300 [36:00<8:48:45, 10.12s/it]2026-04-12 11:22:12,310 | INFO | train_grpo_train | metrics_logged mode=train step=166

                                                    
{'loss': -0.1163, 'grad_norm': 24.627695083618164, 'learning_rate': 9.5e-06, 'num_tokens': 351254.0, 'completions/mean_length': 22.0, 'completions/min_length': 14.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.0, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.4842112064361572, 'rewards/meter/std': 0.49465081095695496, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.8803104162216187, 'rewards/lexical_plausibility/std': 0.18854932487010956, 'rewards/judge_quality/mean': 0.5049999952316284, 'rewards/judge_quality/std': 0.36792856454849243, 'rewards/repeat_penalty/mean': 0.7131342887878418, 'rewards/repeat_penalty/std': 0.07621722668409348, 'rewards/near_duplicate_penalty/mean': 0.9508456587791443, 'rewards/near_duplicate_penalty/std': 0.10162296891212463, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3042754530906677, 'rewards/total_composite/std': 0.4045744836330414, 'reward': 0.3042754530906677, 'reward_std': 0.40457454323768616, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24987587332725525, 'sampling/sampling_logp_difference/max': 1.772230625152588, 'sampling/importance_sampling_ratio/min': 0.16995346546173096, 'sampling/importance_sampling_ratio/mean': 0.9767836332321167, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.409932479262352, 'clip_ratio/low_mean': 0.10937500186264515, 'clip_ratio/low_min': 0.10937500186264515, 'clip_ratio/high_mean': 0.0972222238779068, 'clip_ratio/high_max': 0.0972222238779068, 'clip_ratio/region_mean': 0.20659722574055195, 'reward_total_mean': 0.3042754530906677, 'reward_meter_mean': 0.4842112064361572, 'reward_meter_std': 0.49465081095695496, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.8803104162216187, 'reward_lexical_plausibility_std': 0.18854932487010956, 'reward_repeat_penalty_mean': 0.7131342887878418, 'reward_repeat_penalty_std': 0.07621722668409348, 'reward_near_duplicate_penalty_mean': 0.9508456587791443, 'reward_near_duplicate_penalty_std': 0.10162296891212463, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5049999952316284, 'reward_judge_quality_std': 0.36792856454849243, 'reward_total_composite_mean': 0.3042754530906677, 'reward_total_composite_std': 0.4045744836330414, 'epoch': 0.01}

  5%|▌         | 166/3300 [36:00<8:48:45, 10.12s/it]INFO 04-12 11:22:12 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 167/3300 [36:09<8:38:19,  9.93s/it]2026-04-12 11:22:21,776 | INFO | train_grpo_train | metrics_logged mode=train step=167

                                                    
{'loss': -0.0226, 'grad_norm': 25.09771156311035, 'learning_rate': 9.496969696969698e-06, 'num_tokens': 352590.0, 'completions/mean_length': 20.0, 'completions/min_length': 17.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.0, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.5989440679550171, 'rewards/meter/std': 0.47300752997398376, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.936274528503418, 'rewards/lexical_plausibility/std': 0.11984409391880035, 'rewards/judge_quality/mean': 0.5637499690055847, 'rewards/judge_quality/std': 0.3091896176338196, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.41322803497314453, 'rewards/total_composite/std': 0.3877599239349365, 'reward': 0.41322803497314453, 'reward_std': 0.3877599239349365, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20672015845775604, 'sampling/sampling_logp_difference/max': 2.3006649017333984, 'sampling/importance_sampling_ratio/min': 0.10019221156835556, 'sampling/importance_sampling_ratio/mean': 0.9889694452285767, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1633053570985794, 'clip_ratio/low_mean': 0.07804738450795412, 'clip_ratio/low_min': 0.07804738450795412, 'clip_ratio/high_mean': 0.09561966173350811, 'clip_ratio/high_max': 0.09561966173350811, 'clip_ratio/region_mean': 0.17366704624146223, 'reward_total_mean': 0.41322803497314453, 'reward_meter_mean': 0.5989440679550171, 'reward_meter_std': 0.47300752997398376, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.936274528503418, 'reward_lexical_plausibility_std': 0.11984409391880035, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5637499690055847, 'reward_judge_quality_std': 0.3091896176338196, 'reward_total_composite_mean': 0.41322803497314453, 'reward_total_composite_std': 0.3877599239349365, 'epoch': 0.01}

  5%|▌         | 167/3300 [36:09<8:38:19,  9.93s/it]INFO 04-12 11:22:22 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 168/3300 [36:18<8:23:14,  9.64s/it]2026-04-12 11:22:30,751 | INFO | train_grpo_train | metrics_logged mode=train step=168

                                                    
{'loss': 0.0315, 'grad_norm': 12.210685729980469, 'learning_rate': 9.493939393939395e-06, 'num_tokens': 354230.0, 'completions/mean_length': 58.0, 'completions/min_length': 49.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.0, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.3453054130077362, 'rewards/meter/std': 0.21523037552833557, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9944852590560913, 'rewards/lexical_plausibility/std': 0.015597938559949398, 'rewards/judge_quality/mean': 0.44999998807907104, 'rewards/judge_quality/std': 0.13093073666095734, 'rewards/repeat_penalty/mean': 0.9783644676208496, 'rewards/repeat_penalty/std': 0.031180918216705322, 'rewards/near_duplicate_penalty/mean': 0.9897348284721375, 'rewards/near_duplicate_penalty/std': 0.011738686822354794, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9883342981338501, 'rewards/distinct_2/std': 0.021613944321870804, 'rewards/total_composite/mean': 0.1511130928993225, 'rewards/total_composite/std': 0.08983789384365082, 'reward': 0.1511130928993225, 'reward_std': 0.08983788639307022, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21704605221748352, 'sampling/sampling_logp_difference/max': 1.178271770477295, 'sampling/importance_sampling_ratio/min': 0.3078102469444275, 'sampling/importance_sampling_ratio/mean': 1.0729085206985474, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.014581650495529, 'clip_ratio/low_mean': 0.1251058243215084, 'clip_ratio/low_min': 0.1251058243215084, 'clip_ratio/high_mean': 0.07275232020765543, 'clip_ratio/high_max': 0.07275232020765543, 'clip_ratio/region_mean': 0.19785814452916384, 'reward_total_mean': 0.1511130928993225, 'reward_meter_mean': 0.3453054130077362, 'reward_meter_std': 0.21523037552833557, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9944852590560913, 'reward_lexical_plausibility_std': 0.015597938559949398, 'reward_repeat_penalty_mean': 0.9783644676208496, 'reward_repeat_penalty_std': 0.031180918216705322, 'reward_near_duplicate_penalty_mean': 0.9897348284721375, 'reward_near_duplicate_penalty_std': 0.011738686822354794, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9883342981338501, 'reward_distinct_2_std': 0.021613944321870804, 'reward_judge_quality_mean': 0.44999998807907104, 'reward_judge_quality_std': 0.13093073666095734, 'reward_total_composite_mean': 0.1511130928993225, 'reward_total_composite_std': 0.08983789384365082, 'epoch': 0.01}

  5%|▌         | 168/3300 [36:18<8:23:14,  9.64s/it]INFO 04-12 11:22:31 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 169/3300 [36:27<8:05:15,  9.30s/it]2026-04-12 11:22:39,254 | INFO | train_grpo_train | metrics_logged mode=train step=169

                                                    
{'loss': 0.0824, 'grad_norm': 25.170482635498047, 'learning_rate': 9.490909090909092e-06, 'num_tokens': 355742.0, 'completions/mean_length': 30.0, 'completions/min_length': 25.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.0, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.7387362718582153, 'rewards/meter/std': 0.3245859742164612, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7762500047683716, 'rewards/judge_quality/std': 0.30500292778015137, 'rewards/repeat_penalty/mean': 0.999159038066864, 'rewards/repeat_penalty/std': 0.0021877859253436327, 'rewards/near_duplicate_penalty/mean': 0.999159038066864, 'rewards/near_duplicate_penalty/std': 0.0021877859253436327, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5715652108192444, 'rewards/total_composite/std': 0.4069288671016693, 'reward': 0.5715652108192444, 'reward_std': 0.4069288671016693, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24306611716747284, 'sampling/sampling_logp_difference/max': 2.233613967895508, 'sampling/importance_sampling_ratio/min': 0.10714053362607956, 'sampling/importance_sampling_ratio/mean': 1.0054264068603516, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1133714690804482, 'clip_ratio/low_mean': 0.05826154723763466, 'clip_ratio/low_min': 0.05826154723763466, 'clip_ratio/high_mean': 0.10700204130262136, 'clip_ratio/high_max': 0.10700204130262136, 'clip_ratio/region_mean': 0.16526358854025602, 'reward_total_mean': 0.5715652108192444, 'reward_meter_mean': 0.7387362718582153, 'reward_meter_std': 0.3245859742164612, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.999159038066864, 'reward_repeat_penalty_std': 0.0021877859253436327, 'reward_near_duplicate_penalty_mean': 0.999159038066864, 'reward_near_duplicate_penalty_std': 0.0021877859253436327, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7762500047683716, 'reward_judge_quality_std': 0.30500292778015137, 'reward_total_composite_mean': 0.5715652108192444, 'reward_total_composite_std': 0.4069288671016693, 'epoch': 0.01}

  5%|▌         | 169/3300 [36:27<8:05:15,  9.30s/it]INFO 04-12 11:22:39 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 170/3300 [36:38<8:31:42,  9.81s/it]2026-04-12 11:22:50,253 | INFO | train_grpo_train | metrics_logged mode=train step=170

                                                    
{'loss': 0.0779, 'grad_norm': 9.174952507019043, 'learning_rate': 9.487878787878788e-06, 'num_tokens': 358497.0, 'completions/mean_length': 137.375, 'completions/min_length': 114.0, 'completions/max_length': 199.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 137.375, 'completions/min_terminated_length': 114.0, 'completions/max_terminated_length': 199.0, 'rewards/meter/mean': 0.6350411176681519, 'rewards/meter/std': 0.2673247158527374, 'rewards/count_adherence/mean': 0.9166666269302368, 'rewards/count_adherence/std': 0.0890870913863182, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9683523774147034, 'rewards/lexical_plausibility/std': 0.05798772722482681, 'rewards/judge_quality/mean': 0.21250000596046448, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.9900543689727783, 'rewards/repeat_penalty/std': 0.011428109370172024, 'rewards/near_duplicate_penalty/mean': 0.9962649345397949, 'rewards/near_duplicate_penalty/std': 0.0027726679109036922, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9937705993652344, 'rewards/distinct_2/std': 0.01156657189130783, 'rewards/total_composite/mean': 0.1064797192811966, 'rewards/total_composite/std': 0.07059706747531891, 'reward': 0.1064797192811966, 'reward_std': 0.07059706002473831, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23334676027297974, 'sampling/sampling_logp_difference/max': 1.7164735794067383, 'sampling/importance_sampling_ratio/min': 0.17969873547554016, 'sampling/importance_sampling_ratio/mean': 1.0611952543258667, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.4087454676628113, 'clip_ratio/low_mean': 0.08591734897345304, 'clip_ratio/low_min': 0.08591734897345304, 'clip_ratio/high_mean': 0.10519460588693619, 'clip_ratio/high_max': 0.10519460588693619, 'clip_ratio/region_mean': 0.19111195486038923, 'reward_total_mean': 0.1064797192811966, 'reward_meter_mean': 0.6350411176681519, 'reward_meter_std': 0.2673247158527374, 'reward_count_adherence_mean': 0.9166666269302368, 'reward_count_adherence_std': 0.0890870913863182, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9683523774147034, 'reward_lexical_plausibility_std': 0.05798772722482681, 'reward_repeat_penalty_mean': 0.9900543689727783, 'reward_repeat_penalty_std': 0.011428109370172024, 'reward_near_duplicate_penalty_mean': 0.9962649345397949, 'reward_near_duplicate_penalty_std': 0.0027726679109036922, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9937705993652344, 'reward_distinct_2_std': 0.01156657189130783, 'reward_judge_quality_mean': 0.21250000596046448, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.1064797192811966, 'reward_total_composite_std': 0.07059706747531891, 'epoch': 0.01}

  5%|▌         | 170/3300 [36:38<8:31:42,  9.81s/it]INFO 04-12 11:22:50 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 171/3300 [36:47<8:21:46,  9.62s/it]2026-04-12 11:22:59,438 | INFO | train_grpo_train | metrics_logged mode=train step=171

                                                    
{'loss': 0.0581, 'grad_norm': 20.94089126586914, 'learning_rate': 9.484848484848485e-06, 'num_tokens': 360021.0, 'completions/mean_length': 25.5, 'completions/min_length': 20.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.5, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.3693656325340271, 'rewards/meter/std': 0.3645670413970947, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7475000023841858, 'rewards/judge_quality/std': 0.24656209349632263, 'rewards/repeat_penalty/mean': 0.8291299939155579, 'rewards/repeat_penalty/std': 0.10013485699892044, 'rewards/near_duplicate_penalty/mean': 0.8717963695526123, 'rewards/near_duplicate_penalty/std': 0.08751071989536285, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9516989588737488, 'rewards/distinct_2/std': 0.06690994650125504, 'rewards/total_composite/mean': 0.24836450815200806, 'rewards/total_composite/std': 0.2758057117462158, 'reward': 0.24836450815200806, 'reward_std': 0.27580568194389343, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20083513855934143, 'sampling/sampling_logp_difference/max': 2.0092697143554688, 'sampling/importance_sampling_ratio/min': 0.13408656418323517, 'sampling/importance_sampling_ratio/mean': 0.984898030757904, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.083882451057434, 'clip_ratio/low_mean': 0.11245199851691723, 'clip_ratio/low_min': 0.11245199851691723, 'clip_ratio/high_mean': 0.053392261266708374, 'clip_ratio/high_max': 0.053392261266708374, 'clip_ratio/region_mean': 0.1658442597836256, 'reward_total_mean': 0.24836450815200806, 'reward_meter_mean': 0.3693656325340271, 'reward_meter_std': 0.3645670413970947, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8291299939155579, 'reward_repeat_penalty_std': 0.10013485699892044, 'reward_near_duplicate_penalty_mean': 0.8717963695526123, 'reward_near_duplicate_penalty_std': 0.08751071989536285, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9516989588737488, 'reward_distinct_2_std': 0.06690994650125504, 'reward_judge_quality_mean': 0.7475000023841858, 'reward_judge_quality_std': 0.24656209349632263, 'reward_total_composite_mean': 0.24836450815200806, 'reward_total_composite_std': 0.2758057117462158, 'epoch': 0.01}

  5%|▌         | 171/3300 [36:47<8:21:46,  9.62s/it]INFO 04-12 11:22:59 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 172/3300 [36:56<8:20:13,  9.60s/it]2026-04-12 11:23:08,972 | INFO | train_grpo_train | metrics_logged mode=train step=172

                                                    
{'loss': 0.0456, 'grad_norm': 22.703914642333984, 'learning_rate': 9.481818181818182e-06, 'num_tokens': 361689.0, 'completions/mean_length': 31.5, 'completions/min_length': 26.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.5, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.08846498280763626, 'rewards/meter/std': 0.07687108963727951, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.1060660183429718, 'rewards/repeat_penalty/mean': 0.9858841896057129, 'rewards/repeat_penalty/std': 0.02502317726612091, 'rewards/near_duplicate_penalty/mean': 0.9858841896057129, 'rewards/near_duplicate_penalty/std': 0.02502317726612091, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.0392117016017437, 'rewards/total_composite/std': 0.03517558053135872, 'reward': 0.0392117016017437, 'reward_std': 0.03517558053135872, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22029566764831543, 'sampling/sampling_logp_difference/max': 1.6291742324829102, 'sampling/importance_sampling_ratio/min': 0.196091428399086, 'sampling/importance_sampling_ratio/mean': 1.0678179264068604, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4302738904953003, 'clip_ratio/low_mean': 0.08146736770868301, 'clip_ratio/low_min': 0.08146736770868301, 'clip_ratio/high_mean': 0.07291666977107525, 'clip_ratio/high_max': 0.07291666977107525, 'clip_ratio/region_mean': 0.15438403747975826, 'reward_total_mean': 0.0392117016017437, 'reward_meter_mean': 0.08846498280763626, 'reward_meter_std': 0.07687108963727951, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9858841896057129, 'reward_repeat_penalty_std': 0.02502317726612091, 'reward_near_duplicate_penalty_mean': 0.9858841896057129, 'reward_near_duplicate_penalty_std': 0.02502317726612091, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.1060660183429718, 'reward_total_composite_mean': 0.0392117016017437, 'reward_total_composite_std': 0.03517558053135872, 'epoch': 0.01}

  5%|▌         | 172/3300 [36:56<8:20:13,  9.60s/it]INFO 04-12 11:23:09 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 173/3300 [37:07<8:40:22,  9.98s/it]2026-04-12 11:23:19,866 | INFO | train_grpo_train | metrics_logged mode=train step=173

                                                    
{'loss': 0.0196, 'grad_norm': 8.485453605651855, 'learning_rate': 9.47878787878788e-06, 'num_tokens': 364128.0, 'completions/mean_length': 105.875, 'completions/min_length': 75.0, 'completions/max_length': 140.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 105.875, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 140.0, 'rewards/meter/mean': 0.9131350517272949, 'rewards/meter/std': 0.19232678413391113, '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/lexical_plausibility/mean': 0.9669399857521057, 'rewards/lexical_plausibility/std': 0.06768521666526794, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.10690449178218842, 'rewards/repeat_penalty/mean': 0.9855475425720215, 'rewards/repeat_penalty/std': 0.01551891304552555, 'rewards/near_duplicate_penalty/mean': 0.9928019642829895, 'rewards/near_duplicate_penalty/std': 0.004954707808792591, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9926779270172119, 'rewards/distinct_2/std': 0.013638022355735302, 'rewards/total_composite/mean': 0.2092924267053604, 'rewards/total_composite/std': 0.08456514775753021, 'reward': 0.2092924267053604, 'reward_std': 0.08456514775753021, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2264847457408905, 'sampling/sampling_logp_difference/max': 1.2399978637695312, 'sampling/importance_sampling_ratio/min': 0.2893848419189453, 'sampling/importance_sampling_ratio/mean': 1.0586601495742798, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.2237583100795746, 'clip_ratio/low_mean': 0.11364763416349888, 'clip_ratio/low_min': 0.11364763416349888, 'clip_ratio/high_mean': 0.08463040925562382, 'clip_ratio/high_max': 0.08463040925562382, 'clip_ratio/region_mean': 0.1982780434191227, 'reward_total_mean': 0.2092924267053604, 'reward_meter_mean': 0.9131350517272949, 'reward_meter_std': 0.19232678413391113, '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_lexical_plausibility_mean': 0.9669399857521057, 'reward_lexical_plausibility_std': 0.06768521666526794, 'reward_repeat_penalty_mean': 0.9855475425720215, 'reward_repeat_penalty_std': 0.01551891304552555, 'reward_near_duplicate_penalty_mean': 0.9928019642829895, 'reward_near_duplicate_penalty_std': 0.004954707808792591, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9926779270172119, 'reward_distinct_2_std': 0.013638022355735302, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.10690449178218842, 'reward_total_composite_mean': 0.2092924267053604, 'reward_total_composite_std': 0.08456514775753021, 'epoch': 0.01}

  5%|▌         | 173/3300 [37:07<8:40:22,  9.98s/it]INFO 04-12 11:23:20 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 174/3300 [37:17<8:41:30, 10.01s/it]2026-04-12 11:23:29,933 | INFO | train_grpo_train | metrics_logged mode=train step=174

                                                    
{'loss': 0.0219, 'grad_norm': 15.158143043518066, 'learning_rate': 9.475757575757577e-06, 'num_tokens': 366174.0, 'completions/mean_length': 74.75, 'completions/min_length': 64.0, 'completions/max_length': 104.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.75, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 104.0, 'rewards/meter/mean': 0.7601040601730347, 'rewards/meter/std': 0.3118656575679779, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.528749942779541, 'rewards/judge_quality/std': 0.15824371576309204, 'rewards/repeat_penalty/mean': 0.9680019021034241, 'rewards/repeat_penalty/std': 0.04557621479034424, 'rewards/near_duplicate_penalty/mean': 0.9833802580833435, 'rewards/near_duplicate_penalty/std': 0.016271071508526802, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9839502573013306, 'rewards/distinct_2/std': 0.03230852633714676, 'rewards/total_composite/mean': 0.33249223232269287, 'rewards/total_composite/std': 0.21956424415111542, 'reward': 0.33249223232269287, 'reward_std': 0.2195642590522766, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21080315113067627, 'sampling/sampling_logp_difference/max': 1.8717031478881836, 'sampling/importance_sampling_ratio/min': 0.1538613885641098, 'sampling/importance_sampling_ratio/mean': 1.0346046686172485, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0482437312602997, 'clip_ratio/low_mean': 0.061938654631376266, 'clip_ratio/low_min': 0.061938654631376266, 'clip_ratio/high_mean': 0.13117137737572193, 'clip_ratio/high_max': 0.13117137737572193, 'clip_ratio/region_mean': 0.1931100320070982, 'reward_total_mean': 0.33249223232269287, 'reward_meter_mean': 0.7601040601730347, 'reward_meter_std': 0.3118656575679779, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9680019021034241, 'reward_repeat_penalty_std': 0.04557621479034424, 'reward_near_duplicate_penalty_mean': 0.9833802580833435, 'reward_near_duplicate_penalty_std': 0.016271071508526802, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9839502573013306, 'reward_distinct_2_std': 0.03230852633714676, 'reward_judge_quality_mean': 0.528749942779541, 'reward_judge_quality_std': 0.15824371576309204, 'reward_total_composite_mean': 0.33249223232269287, 'reward_total_composite_std': 0.21956424415111542, 'epoch': 0.01}

  5%|▌         | 174/3300 [37:17<8:41:30, 10.01s/it]INFO 04-12 11:23:30 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 175/3300 [37:26<8:27:04,  9.74s/it]2026-04-12 11:23:39,029 | INFO | train_grpo_train | metrics_logged mode=train step=175

                                                    
{'loss': -0.0079, 'grad_norm': 18.586265563964844, 'learning_rate': 9.472727272727274e-06, 'num_tokens': 367733.0, 'completions/mean_length': 38.875, 'completions/min_length': 31.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.875, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.5412830114364624, 'rewards/meter/std': 0.41530048847198486, '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/lexical_plausibility/mean': 0.9761029481887817, 'rewards/lexical_plausibility/std': 0.067591093480587, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.10690449178218842, 'rewards/repeat_penalty/mean': 0.9654839038848877, 'rewards/repeat_penalty/std': 0.025515692308545113, 'rewards/near_duplicate_penalty/mean': 0.9654839038848877, 'rewards/near_duplicate_penalty/std': 0.025515692308545113, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.19817006587982178, 'rewards/total_composite/std': 0.18958911299705505, 'reward': 0.19817006587982178, 'reward_std': 0.18958912789821625, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2097066193819046, 'sampling/sampling_logp_difference/max': 1.918522834777832, 'sampling/importance_sampling_ratio/min': 0.1468236893415451, 'sampling/importance_sampling_ratio/mean': 1.0218008756637573, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6516703888773918, 'clip_ratio/low_mean': 0.10428987257182598, 'clip_ratio/low_min': 0.10428987257182598, 'clip_ratio/high_mean': 0.12269819341599941, 'clip_ratio/high_max': 0.12269819341599941, 'clip_ratio/region_mean': 0.2269880659878254, 'reward_total_mean': 0.19817006587982178, 'reward_meter_mean': 0.5412830114364624, 'reward_meter_std': 0.41530048847198486, '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_lexical_plausibility_mean': 0.9761029481887817, 'reward_lexical_plausibility_std': 0.067591093480587, 'reward_repeat_penalty_mean': 0.9654839038848877, 'reward_repeat_penalty_std': 0.025515692308545113, 'reward_near_duplicate_penalty_mean': 0.9654839038848877, 'reward_near_duplicate_penalty_std': 0.025515692308545113, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.10690449178218842, 'reward_total_composite_mean': 0.19817006587982178, 'reward_total_composite_std': 0.18958911299705505, 'epoch': 0.01}

  5%|▌         | 175/3300 [37:27<8:27:04,  9.74s/it]INFO 04-12 11:23:39 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 176/3300 [37:40<9:24:08, 10.83s/it]2026-04-12 11:23:52,430 | INFO | train_grpo_train | metrics_logged mode=train step=176

                                                    
{'loss': -0.0022, 'grad_norm': 5.116064548492432, 'learning_rate': 9.469696969696971e-06, 'num_tokens': 371605.0, 'completions/mean_length': 277.0, 'completions/min_length': 253.0, 'completions/max_length': 325.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 277.0, 'completions/min_terminated_length': 253.0, 'completions/max_terminated_length': 325.0, 'rewards/meter/mean': 0.6602240800857544, 'rewards/meter/std': 0.23572297394275665, 'rewards/count_adherence/mean': 0.9230769276618958, 'rewards/count_adherence/std': 0.07121692597866058, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9989809989929199, 'rewards/lexical_plausibility/std': 0.00288223335519433, 'rewards/judge_quality/mean': 0.20000000298023224, 'rewards/judge_quality/std': 0.07559289038181305, 'rewards/repeat_penalty/mean': 0.980717122554779, 'rewards/repeat_penalty/std': 0.018925588577985764, 'rewards/near_duplicate_penalty/mean': 0.9932653903961182, 'rewards/near_duplicate_penalty/std': 0.0036780915688723326, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.987325131893158, 'rewards/distinct_2/std': 0.016033917665481567, 'rewards/total_composite/mean': 0.11684560030698776, 'rewards/total_composite/std': 0.08322425931692123, 'reward': 0.11684560030698776, 'reward_std': 0.08322425931692123, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.235348641872406, 'sampling/sampling_logp_difference/max': 2.1670923233032227, 'sampling/importance_sampling_ratio/min': 0.11451008915901184, 'sampling/importance_sampling_ratio/mean': 1.0630502700805664, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 4.109456360340118, 'clip_ratio/low_mean': 0.08097132854163647, 'clip_ratio/low_min': 0.08097132854163647, 'clip_ratio/high_mean': 0.10114516504108906, 'clip_ratio/high_max': 0.10114516504108906, 'clip_ratio/region_mean': 0.18211649358272552, 'reward_total_mean': 0.11684560030698776, 'reward_meter_mean': 0.6602240800857544, 'reward_meter_std': 0.23572297394275665, 'reward_count_adherence_mean': 0.9230769276618958, 'reward_count_adherence_std': 0.07121692597866058, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9989809989929199, 'reward_lexical_plausibility_std': 0.00288223335519433, 'reward_repeat_penalty_mean': 0.980717122554779, 'reward_repeat_penalty_std': 0.018925588577985764, 'reward_near_duplicate_penalty_mean': 0.9932653903961182, 'reward_near_duplicate_penalty_std': 0.0036780915688723326, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.987325131893158, 'reward_distinct_2_std': 0.016033917665481567, 'reward_judge_quality_mean': 0.20000000298023224, 'reward_judge_quality_std': 0.07559289038181305, 'reward_total_composite_mean': 0.11684560030698776, 'reward_total_composite_std': 0.08322425931692123, 'epoch': 0.01}

  5%|▌         | 176/3300 [37:40<9:24:08, 10.83s/it]INFO 04-12 11:23:52 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 177/3300 [37:54<10:14:57, 11.81s/it]2026-04-12 11:24:06,531 | INFO | train_grpo_train | metrics_logged mode=train step=177

                                                     
{'loss': 0.1403, 'grad_norm': 6.402902603149414, 'learning_rate': 9.466666666666667e-06, 'num_tokens': 375200.0, 'completions/mean_length': 233.375, 'completions/min_length': 161.0, 'completions/max_length': 359.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 233.375, 'completions/min_terminated_length': 161.0, 'completions/max_terminated_length': 359.0, 'rewards/meter/mean': 0.4629499018192291, 'rewards/meter/std': 0.28422310948371887, 'rewards/count_adherence/mean': 0.824999988079071, 'rewards/count_adherence/std': 0.11649646610021591, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9468951225280762, 'rewards/lexical_plausibility/std': 0.06595577299594879, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.996282696723938, 'rewards/repeat_penalty/std': 0.003457123413681984, 'rewards/near_duplicate_penalty/mean': 0.9974746108055115, 'rewards/near_duplicate_penalty/std': 0.0012659269850701094, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9988055229187012, 'rewards/distinct_2/std': 0.0033784282859414816, 'rewards/total_composite/mean': 0.055360257625579834, 'rewards/total_composite/std': 0.04187962785363197, 'reward': 0.055360257625579834, 'reward_std': 0.04187962785363197, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23183801770210266, 'sampling/sampling_logp_difference/max': 1.3596258163452148, 'sampling/importance_sampling_ratio/min': 0.25675684213638306, 'sampling/importance_sampling_ratio/mean': 1.0723868608474731, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 4.344330847263336, 'clip_ratio/low_mean': 0.08006827719509602, 'clip_ratio/low_min': 0.08006827719509602, 'clip_ratio/high_mean': 0.09986113384366035, 'clip_ratio/high_max': 0.09986113384366035, 'clip_ratio/region_mean': 0.17992941103875637, 'reward_total_mean': 0.055360257625579834, 'reward_meter_mean': 0.4629499018192291, 'reward_meter_std': 0.28422310948371887, 'reward_count_adherence_mean': 0.824999988079071, 'reward_count_adherence_std': 0.11649646610021591, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9468951225280762, 'reward_lexical_plausibility_std': 0.06595577299594879, 'reward_repeat_penalty_mean': 0.996282696723938, 'reward_repeat_penalty_std': 0.003457123413681984, 'reward_near_duplicate_penalty_mean': 0.9974746108055115, 'reward_near_duplicate_penalty_std': 0.0012659269850701094, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9988055229187012, 'reward_distinct_2_std': 0.0033784282859414816, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.055360257625579834, 'reward_total_composite_std': 0.04187962785363197, 'epoch': 0.01}

  5%|▌         | 177/3300 [37:54<10:14:57, 11.81s/it]INFO 04-12 11:24:06 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 178/3300 [38:03<9:37:01, 11.09s/it] 2026-04-12 11:24:15,928 | INFO | train_grpo_train | metrics_logged mode=train step=178

                                                    
{'loss': 0.0505, 'grad_norm': 14.922652244567871, 'learning_rate': 9.463636363636364e-06, 'num_tokens': 376912.0, 'completions/mean_length': 43.0, 'completions/min_length': 34.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.0, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.7356902956962585, 'rewards/meter/std': 0.37158194184303284, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9743161201477051, 'rewards/lexical_plausibility/std': 0.05866875872015953, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.9899815917015076, 'rewards/repeat_penalty/std': 0.025666886940598488, 'rewards/near_duplicate_penalty/mean': 0.999139130115509, 'rewards/near_duplicate_penalty/std': 0.002434949856251478, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9908424615859985, 'rewards/distinct_2/std': 0.025901345536112785, 'rewards/total_composite/mean': 0.2536138892173767, 'rewards/total_composite/std': 0.17918303608894348, 'reward': 0.2536138892173767, 'reward_std': 0.17918303608894348, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22188541293144226, 'sampling/sampling_logp_difference/max': 1.4895896911621094, 'sampling/importance_sampling_ratio/min': 0.22546514868736267, 'sampling/importance_sampling_ratio/mean': 1.0171048641204834, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2047335356473923, 'clip_ratio/low_mean': 0.1056149760261178, 'clip_ratio/low_min': 0.1056149760261178, 'clip_ratio/high_mean': 0.12168767675757408, 'clip_ratio/high_max': 0.12168767675757408, 'clip_ratio/region_mean': 0.22730265278369188, 'reward_total_mean': 0.2536138892173767, 'reward_meter_mean': 0.7356902956962585, 'reward_meter_std': 0.37158194184303284, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9743161201477051, 'reward_lexical_plausibility_std': 0.05866875872015953, 'reward_repeat_penalty_mean': 0.9899815917015076, 'reward_repeat_penalty_std': 0.025666886940598488, 'reward_near_duplicate_penalty_mean': 0.999139130115509, 'reward_near_duplicate_penalty_std': 0.002434949856251478, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9908424615859985, 'reward_distinct_2_std': 0.025901345536112785, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.2536138892173767, 'reward_total_composite_std': 0.17918303608894348, 'epoch': 0.01}

  5%|▌         | 178/3300 [38:03<9:37:01, 11.09s/it]INFO 04-12 11:24:16 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 179/3300 [38:13<9:14:51, 10.67s/it]2026-04-12 11:24:25,609 | INFO | train_grpo_train | metrics_logged mode=train step=179

                                                    
{'loss': 0.0648, 'grad_norm': 16.850215911865234, 'learning_rate': 9.460606060606061e-06, 'num_tokens': 378408.0, 'completions/mean_length': 38.0, 'completions/min_length': 33.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.0, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.41254815459251404, 'rewards/meter/std': 0.44692495465278625, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9485032558441162, 'rewards/lexical_plausibility/std': 0.08522287011146545, 'rewards/judge_quality/mean': 0.4124999940395355, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.9899126291275024, 'rewards/repeat_penalty/std': 0.02286543883383274, 'rewards/near_duplicate_penalty/mean': 0.9899126291275024, 'rewards/near_duplicate_penalty/std': 0.02286543883383274, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.16204196214675903, 'rewards/total_composite/std': 0.19023385643959045, 'reward': 0.16204196214675903, 'reward_std': 0.19023385643959045, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24171759188175201, 'sampling/sampling_logp_difference/max': 1.396334171295166, 'sampling/importance_sampling_ratio/min': 0.24750261008739471, 'sampling/importance_sampling_ratio/mean': 1.0306048393249512, 'sampling/importance_sampling_ratio/max': 1.823089838027954, 'entropy': 2.6068559288978577, 'clip_ratio/low_mean': 0.11815614067018032, 'clip_ratio/low_min': 0.11815614067018032, 'clip_ratio/high_mean': 0.06796252168715, 'clip_ratio/high_max': 0.06796252168715, 'clip_ratio/region_mean': 0.18611866235733032, 'reward_total_mean': 0.16204196214675903, 'reward_meter_mean': 0.41254815459251404, 'reward_meter_std': 0.44692495465278625, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9485032558441162, 'reward_lexical_plausibility_std': 0.08522287011146545, 'reward_repeat_penalty_mean': 0.9899126291275024, 'reward_repeat_penalty_std': 0.02286543883383274, 'reward_near_duplicate_penalty_mean': 0.9899126291275024, 'reward_near_duplicate_penalty_std': 0.02286543883383274, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4124999940395355, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.16204196214675903, 'reward_total_composite_std': 0.19023385643959045, 'epoch': 0.01}

  5%|▌         | 179/3300 [38:13<9:14:51, 10.67s/it]INFO 04-12 11:24:25 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 180/3300 [38:22<8:53:06, 10.25s/it]2026-04-12 11:24:34,892 | INFO | train_grpo_train | metrics_logged mode=train step=180

                                                    
{'loss': 0.1504, 'grad_norm': 19.240222930908203, 'learning_rate': 9.457575757575759e-06, 'num_tokens': 380440.0, 'completions/mean_length': 71.0, 'completions/min_length': 55.0, 'completions/max_length': 85.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.0, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 85.0, 'rewards/meter/mean': 0.42972785234451294, 'rewards/meter/std': 0.28869742155075073, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9831730723381042, 'rewards/lexical_plausibility/std': 0.04759372025728226, 'rewards/judge_quality/mean': 0.5049999952316284, 'rewards/judge_quality/std': 0.2745385766029358, 'rewards/repeat_penalty/mean': 0.9570775032043457, 'rewards/repeat_penalty/std': 0.06302423775196075, 'rewards/near_duplicate_penalty/mean': 0.9806978702545166, 'rewards/near_duplicate_penalty/std': 0.01663300208747387, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9753398299217224, 'rewards/distinct_2/std': 0.05175751447677612, 'rewards/total_composite/mean': 0.1683695912361145, 'rewards/total_composite/std': 0.10020577907562256, 'reward': 0.1683695912361145, 'reward_std': 0.10020577907562256, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23795141279697418, 'sampling/sampling_logp_difference/max': 2.5275683403015137, 'sampling/importance_sampling_ratio/min': 0.07985296100378036, 'sampling/importance_sampling_ratio/mean': 1.0296127796173096, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8202795684337616, 'clip_ratio/low_mean': 0.09395610727369785, 'clip_ratio/low_min': 0.09395610727369785, 'clip_ratio/high_mean': 0.12110170815140009, 'clip_ratio/high_max': 0.12110170815140009, 'clip_ratio/region_mean': 0.21505781542509794, 'reward_total_mean': 0.1683695912361145, 'reward_meter_mean': 0.42972785234451294, 'reward_meter_std': 0.28869742155075073, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9831730723381042, 'reward_lexical_plausibility_std': 0.04759372025728226, 'reward_repeat_penalty_mean': 0.9570775032043457, 'reward_repeat_penalty_std': 0.06302423775196075, 'reward_near_duplicate_penalty_mean': 0.9806978702545166, 'reward_near_duplicate_penalty_std': 0.01663300208747387, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9753398299217224, 'reward_distinct_2_std': 0.05175751447677612, 'reward_judge_quality_mean': 0.5049999952316284, 'reward_judge_quality_std': 0.2745385766029358, 'reward_total_composite_mean': 0.1683695912361145, 'reward_total_composite_std': 0.10020577907562256, 'epoch': 0.01}

  5%|▌         | 180/3300 [38:22<8:53:06, 10.25s/it]INFO 04-12 11:24:35 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 181/3300 [38:31<8:29:00,  9.79s/it]2026-04-12 11:24:43,610 | INFO | train_grpo_train | metrics_logged mode=train step=181

                                                    
{'loss': -0.0253, 'grad_norm': 17.628551483154297, 'learning_rate': 9.454545454545456e-06, 'num_tokens': 382191.0, 'completions/mean_length': 43.875, 'completions/min_length': 31.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.875, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.2817838191986084, 'rewards/meter/std': 0.34548836946487427, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.17817415297031403, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6725000143051147, 'rewards/judge_quality/std': 0.2666056752204895, 'rewards/repeat_penalty/mean': 0.9958004355430603, 'rewards/repeat_penalty/std': 0.004529349971562624, 'rewards/near_duplicate_penalty/mean': 0.9958004355430603, 'rewards/near_duplicate_penalty/std': 0.004529349971562624, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.12685979902744293, 'rewards/total_composite/std': 0.13998359441757202, 'reward': 0.12685979902744293, 'reward_std': 0.13998357951641083, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1864061951637268, 'sampling/sampling_logp_difference/max': 1.9987711906433105, 'sampling/importance_sampling_ratio/min': 0.1355016976594925, 'sampling/importance_sampling_ratio/mean': 1.0109132528305054, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5137614160776138, 'clip_ratio/low_mean': 0.09676664415746927, 'clip_ratio/low_min': 0.09676664415746927, 'clip_ratio/high_mean': 0.05401897616684437, 'clip_ratio/high_max': 0.05401897616684437, 'clip_ratio/region_mean': 0.15078562032431364, 'reward_total_mean': 0.12685979902744293, 'reward_meter_mean': 0.2817838191986084, 'reward_meter_std': 0.34548836946487427, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.17817415297031403, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9958004355430603, 'reward_repeat_penalty_std': 0.004529349971562624, 'reward_near_duplicate_penalty_mean': 0.9958004355430603, 'reward_near_duplicate_penalty_std': 0.004529349971562624, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6725000143051147, 'reward_judge_quality_std': 0.2666056752204895, 'reward_total_composite_mean': 0.12685979902744293, 'reward_total_composite_std': 0.13998359441757202, 'epoch': 0.01}

  5%|▌         | 181/3300 [38:31<8:29:00,  9.79s/it]INFO 04-12 11:24:43 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 182/3300 [38:41<8:32:58,  9.87s/it]2026-04-12 11:24:53,668 | INFO | train_grpo_train | metrics_logged mode=train step=182

                                                    
{'loss': 0.0259, 'grad_norm': 8.540081977844238, 'learning_rate': 9.451515151515153e-06, 'num_tokens': 384769.0, 'completions/mean_length': 118.25, 'completions/min_length': 95.0, 'completions/max_length': 164.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 118.25, 'completions/min_terminated_length': 95.0, 'completions/max_terminated_length': 164.0, 'rewards/meter/mean': 0.4592788517475128, 'rewards/meter/std': 0.267805814743042, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.9770813584327698, 'rewards/repeat_penalty/std': 0.033471085131168365, 'rewards/near_duplicate_penalty/mean': 0.9907737970352173, 'rewards/near_duplicate_penalty/std': 0.010123512707650661, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9860093593597412, 'rewards/distinct_2/std': 0.025911342352628708, 'rewards/total_composite/mean': 0.12077900767326355, 'rewards/total_composite/std': 0.11277022957801819, 'reward': 0.12077900767326355, 'reward_std': 0.11277022957801819, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22959047555923462, 'sampling/sampling_logp_difference/max': 2.329220771789551, 'sampling/importance_sampling_ratio/min': 0.09737159311771393, 'sampling/importance_sampling_ratio/mean': 1.045446753501892, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.243285685777664, 'clip_ratio/low_mean': 0.09601961821317673, 'clip_ratio/low_min': 0.09601961821317673, 'clip_ratio/high_mean': 0.07793585769832134, 'clip_ratio/high_max': 0.07793585769832134, 'clip_ratio/region_mean': 0.17395547591149807, 'reward_total_mean': 0.12077900767326355, 'reward_meter_mean': 0.4592788517475128, 'reward_meter_std': 0.267805814743042, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.0707106739282608, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.9770813584327698, 'reward_repeat_penalty_std': 0.033471085131168365, 'reward_near_duplicate_penalty_mean': 0.9907737970352173, 'reward_near_duplicate_penalty_std': 0.010123512707650661, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9860093593597412, 'reward_distinct_2_std': 0.025911342352628708, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.12077900767326355, 'reward_total_composite_std': 0.11277022957801819, 'epoch': 0.01}

  6%|▌         | 182/3300 [38:41<8:32:58,  9.87s/it]INFO 04-12 11:24:53 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 183/3300 [38:53<9:00:16, 10.40s/it]2026-04-12 11:25:05,301 | INFO | train_grpo_train | metrics_logged mode=train step=183

                                                    
{'loss': 0.0137, 'grad_norm': 6.802778244018555, 'learning_rate': 9.448484848484849e-06, 'num_tokens': 387736.0, 'completions/mean_length': 169.875, 'completions/min_length': 145.0, 'completions/max_length': 203.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 169.875, 'completions/min_terminated_length': 145.0, 'completions/max_terminated_length': 203.0, 'rewards/meter/mean': 0.6901034116744995, 'rewards/meter/std': 0.2485126554965973, 'rewards/count_adherence/mean': 0.8928571939468384, 'rewards/count_adherence/std': 0.06613000482320786, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9692424535751343, 'rewards/lexical_plausibility/std': 0.04459897801280022, 'rewards/judge_quality/mean': 0.17500001192092896, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/repeat_penalty/mean': 0.9947319030761719, 'rewards/repeat_penalty/std': 0.009320467710494995, 'rewards/near_duplicate_penalty/mean': 0.9968446493148804, 'rewards/near_duplicate_penalty/std': 0.0034918736200779676, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.997863233089447, 'rewards/distinct_2/std': 0.006043647415935993, 'rewards/total_composite/mean': 0.08977998793125153, 'rewards/total_composite/std': 0.07316060364246368, 'reward': 0.08977998793125153, 'reward_std': 0.07316060364246368, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.251144677400589, 'sampling/sampling_logp_difference/max': 1.471531867980957, 'sampling/importance_sampling_ratio/min': 0.22957353293895721, 'sampling/importance_sampling_ratio/mean': 1.0808608531951904, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 4.208578437566757, 'clip_ratio/low_mean': 0.08524911105632782, 'clip_ratio/low_min': 0.08524911105632782, 'clip_ratio/high_mean': 0.09699354693293571, 'clip_ratio/high_max': 0.09699354693293571, 'clip_ratio/region_mean': 0.18224265798926353, 'reward_total_mean': 0.08977998793125153, 'reward_meter_mean': 0.6901034116744995, 'reward_meter_std': 0.2485126554965973, 'reward_count_adherence_mean': 0.8928571939468384, 'reward_count_adherence_std': 0.06613000482320786, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9692424535751343, 'reward_lexical_plausibility_std': 0.04459897801280022, 'reward_repeat_penalty_mean': 0.9947319030761719, 'reward_repeat_penalty_std': 0.009320467710494995, 'reward_near_duplicate_penalty_mean': 0.9968446493148804, 'reward_near_duplicate_penalty_std': 0.0034918736200779676, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.997863233089447, 'reward_distinct_2_std': 0.006043647415935993, 'reward_judge_quality_mean': 0.17500001192092896, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.08977998793125153, 'reward_total_composite_std': 0.07316060364246368, 'epoch': 0.01}

  6%|▌         | 183/3300 [38:53<9:00:16, 10.40s/it]INFO 04-12 11:25:05 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 184/3300 [39:03<8:54:50, 10.30s/it]2026-04-12 11:25:15,362 | INFO | train_grpo_train | metrics_logged mode=train step=184

                                                    
{'loss': 0.0123, 'grad_norm': 13.832329750061035, 'learning_rate': 9.445454545454546e-06, 'num_tokens': 389307.0, 'completions/mean_length': 39.375, 'completions/min_length': 32.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.375, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.7828742265701294, 'rewards/meter/std': 0.3550904393196106, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7549999952316284, 'rewards/judge_quality/std': 0.25286927819252014, 'rewards/repeat_penalty/mean': 0.9965667724609375, 'rewards/repeat_penalty/std': 0.006824270356446505, 'rewards/near_duplicate_penalty/mean': 0.9965667724609375, 'rewards/near_duplicate_penalty/std': 0.006824270356446505, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5553215742111206, 'rewards/total_composite/std': 0.3137626647949219, 'reward': 0.5553215742111206, 'reward_std': 0.31376269459724426, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19523799419403076, 'sampling/sampling_logp_difference/max': 1.07771635055542, 'sampling/importance_sampling_ratio/min': 0.34545567631721497, 'sampling/importance_sampling_ratio/mean': 1.0603926181793213, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.899738535284996, 'clip_ratio/low_mean': 0.1065289827529341, 'clip_ratio/low_min': 0.1065289827529341, 'clip_ratio/high_mean': 0.0513355010189116, 'clip_ratio/high_max': 0.0513355010189116, 'clip_ratio/region_mean': 0.1578644837718457, 'reward_total_mean': 0.5553215742111206, 'reward_meter_mean': 0.7828742265701294, 'reward_meter_std': 0.3550904393196106, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9965667724609375, 'reward_repeat_penalty_std': 0.006824270356446505, 'reward_near_duplicate_penalty_mean': 0.9965667724609375, 'reward_near_duplicate_penalty_std': 0.006824270356446505, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7549999952316284, 'reward_judge_quality_std': 0.25286927819252014, 'reward_total_composite_mean': 0.5553215742111206, 'reward_total_composite_std': 0.3137626647949219, 'epoch': 0.01}

  6%|▌         | 184/3300 [39:03<8:54:50, 10.30s/it]INFO 04-12 11:25:15 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 185/3300 [39:11<8:19:50,  9.63s/it]2026-04-12 11:25:23,426 | INFO | train_grpo_train | metrics_logged mode=train step=185

                                                    
{'loss': 0.0331, 'grad_norm': 21.2224063873291, 'learning_rate': 9.442424242424243e-06, 'num_tokens': 390703.0, 'completions/mean_length': 19.5, 'completions/min_length': 10.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.5, 'completions/min_terminated_length': 10.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.015161988325417042, 'rewards/meter/std': 0.03626592084765434, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.5345224738121033, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.47486287355422974, 'rewards/lexical_plausibility/std': 0.29278481006622314, 'rewards/judge_quality/mean': 0.16249999403953552, 'rewards/judge_quality/std': 0.07905694097280502, 'rewards/repeat_penalty/mean': 0.8109553456306458, 'rewards/repeat_penalty/std': 0.26090797781944275, 'rewards/near_duplicate_penalty/mean': 0.9229865074157715, 'rewards/near_duplicate_penalty/std': 0.2145070880651474, 'rewards/opening_diversity/mean': 0.875, 'rewards/opening_diversity/std': 0.13363061845302582, 'rewards/distinct_2/mean': 0.9786324501037598, 'rewards/distinct_2/std': 0.06043647602200508, 'rewards/total_composite/mean': 0.002280371030792594, 'rewards/total_composite/std': 0.005437059327960014, 'reward': 0.002280371030792594, 'reward_std': 0.005437058862298727, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18426287174224854, 'sampling/sampling_logp_difference/max': 1.4179484844207764, 'sampling/importance_sampling_ratio/min': 0.24221040308475494, 'sampling/importance_sampling_ratio/mean': 1.037623643875122, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.565277323126793, 'clip_ratio/low_mean': 0.13158263638615608, 'clip_ratio/low_min': 0.13158263638615608, 'clip_ratio/high_mean': 0.007352941203862429, 'clip_ratio/high_max': 0.007352941203862429, 'clip_ratio/region_mean': 0.1389355775900185, 'reward_total_mean': 0.002280371030792594, 'reward_meter_mean': 0.015161988325417042, 'reward_meter_std': 0.03626592084765434, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.5345224738121033, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.47486287355422974, 'reward_lexical_plausibility_std': 0.29278481006622314, 'reward_repeat_penalty_mean': 0.8109553456306458, 'reward_repeat_penalty_std': 0.26090797781944275, 'reward_near_duplicate_penalty_mean': 0.9229865074157715, 'reward_near_duplicate_penalty_std': 0.2145070880651474, 'reward_opening_diversity_mean': 0.875, 'reward_opening_diversity_std': 0.13363061845302582, 'reward_distinct_2_mean': 0.9786324501037598, 'reward_distinct_2_std': 0.06043647602200508, 'reward_judge_quality_mean': 0.16249999403953552, 'reward_judge_quality_std': 0.07905694097280502, 'reward_total_composite_mean': 0.002280371030792594, 'reward_total_composite_std': 0.005437059327960014, 'epoch': 0.01}

  6%|▌         | 185/3300 [39:11<8:19:50,  9.63s/it]INFO 04-12 11:25:23 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 186/3300 [39:23<9:03:58, 10.48s/it]2026-04-12 11:25:35,898 | INFO | train_grpo_train | metrics_logged mode=train step=186

                                                    
{'loss': 0.0882, 'grad_norm': 9.80618953704834, 'learning_rate': 9.43939393939394e-06, 'num_tokens': 393278.0, 'completions/mean_length': 136.875, 'completions/min_length': 118.0, 'completions/max_length': 160.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 136.875, 'completions/min_terminated_length': 118.0, 'completions/max_terminated_length': 160.0, 'rewards/meter/mean': 0.5913933515548706, 'rewards/meter/std': 0.39986154437065125, 'rewards/count_adherence/mean': 0.8928571939468384, 'rewards/count_adherence/std': 0.06613000482320786, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9899776577949524, 'rewards/lexical_plausibility/std': 0.02834748476743698, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.9453206062316895, 'rewards/repeat_penalty/std': 0.10422060638666153, 'rewards/near_duplicate_penalty/mean': 0.9900493621826172, 'rewards/near_duplicate_penalty/std': 0.01023941021412611, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.96144700050354, 'rewards/distinct_2/std': 0.07747188210487366, 'rewards/total_composite/mean': 0.1846718043088913, 'rewards/total_composite/std': 0.12800109386444092, 'reward': 0.1846718043088913, 'reward_std': 0.12800109386444092, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23754943907260895, 'sampling/sampling_logp_difference/max': 1.767012357711792, 'sampling/importance_sampling_ratio/min': 0.17084264755249023, 'sampling/importance_sampling_ratio/mean': 1.0489275455474854, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.7036547362804413, 'clip_ratio/low_mean': 0.09044165164232254, 'clip_ratio/low_min': 0.09044165164232254, 'clip_ratio/high_mean': 0.09846928715705872, 'clip_ratio/high_max': 0.09846928715705872, 'clip_ratio/region_mean': 0.18891093879938126, 'reward_total_mean': 0.1846718043088913, 'reward_meter_mean': 0.5913933515548706, 'reward_meter_std': 0.39986154437065125, 'reward_count_adherence_mean': 0.8928571939468384, 'reward_count_adherence_std': 0.06613000482320786, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9899776577949524, 'reward_lexical_plausibility_std': 0.02834748476743698, 'reward_repeat_penalty_mean': 0.9453206062316895, 'reward_repeat_penalty_std': 0.10422060638666153, 'reward_near_duplicate_penalty_mean': 0.9900493621826172, 'reward_near_duplicate_penalty_std': 0.01023941021412611, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.96144700050354, 'reward_distinct_2_std': 0.07747188210487366, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.1846718043088913, 'reward_total_composite_std': 0.12800109386444092, 'epoch': 0.01}

  6%|▌         | 186/3300 [39:23<9:03:58, 10.48s/it]INFO 04-12 11:25:36 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 187/3300 [39:32<8:38:30,  9.99s/it]2026-04-12 11:25:44,754 | INFO | train_grpo_train | metrics_logged mode=train step=187

                                                    
{'loss': -0.0556, 'grad_norm': 16.716724395751953, 'learning_rate': 9.436363636363636e-06, 'num_tokens': 394835.0, 'completions/mean_length': 36.625, 'completions/min_length': 28.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.625, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.4389100670814514, 'rewards/meter/std': 0.3943783938884735, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.47499996423721313, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/repeat_penalty/mean': 0.9256715774536133, 'rewards/repeat_penalty/std': 0.16142699122428894, 'rewards/near_duplicate_penalty/mean': 0.9672102928161621, 'rewards/near_duplicate_penalty/std': 0.04706943780183792, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9773755669593811, 'rewards/distinct_2/std': 0.06399156153202057, 'rewards/total_composite/mean': 0.19750946760177612, 'rewards/total_composite/std': 0.1774703711271286, 'reward': 0.19750946760177612, 'reward_std': 0.1774703711271286, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19253911077976227, 'sampling/sampling_logp_difference/max': 2.2758216857910156, 'sampling/importance_sampling_ratio/min': 0.1027124747633934, 'sampling/importance_sampling_ratio/mean': 1.0290716886520386, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8666294813156128, 'clip_ratio/low_mean': 0.08704298455268145, 'clip_ratio/low_min': 0.08704298455268145, 'clip_ratio/high_mean': 0.051872119307518005, 'clip_ratio/high_max': 0.051872119307518005, 'clip_ratio/region_mean': 0.13891510386019945, 'reward_total_mean': 0.19750946760177612, 'reward_meter_mean': 0.4389100670814514, 'reward_meter_std': 0.3943783938884735, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9256715774536133, 'reward_repeat_penalty_std': 0.16142699122428894, 'reward_near_duplicate_penalty_mean': 0.9672102928161621, 'reward_near_duplicate_penalty_std': 0.04706943780183792, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9773755669593811, 'reward_distinct_2_std': 0.06399156153202057, 'reward_judge_quality_mean': 0.47499996423721313, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.19750946760177612, 'reward_total_composite_std': 0.1774703711271286, 'epoch': 0.01}

  6%|▌         | 187/3300 [39:32<8:38:30,  9.99s/it]INFO 04-12 11:25:44 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 188/3300 [39:42<8:29:28,  9.82s/it]2026-04-12 11:25:54,179 | INFO | train_grpo_train | metrics_logged mode=train step=188

                                                    
{'loss': -0.0403, 'grad_norm': 11.473904609680176, 'learning_rate': 9.433333333333335e-06, 'num_tokens': 396671.0, 'completions/mean_length': 68.5, 'completions/min_length': 50.0, 'completions/max_length': 83.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.5, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.48370596766471863, 'rewards/meter/std': 0.41808390617370605, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9972826242446899, 'rewards/lexical_plausibility/std': 0.007685941644012928, 'rewards/judge_quality/mean': 0.44999998807907104, 'rewards/judge_quality/std': 0.09258200973272324, 'rewards/repeat_penalty/mean': 0.9845765829086304, 'rewards/repeat_penalty/std': 0.01943829655647278, 'rewards/near_duplicate_penalty/mean': 0.9938940405845642, 'rewards/near_duplicate_penalty/std': 0.003942488227039576, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9905967712402344, 'rewards/distinct_2/std': 0.017439009621739388, 'rewards/total_composite/mean': 0.23267492651939392, 'rewards/total_composite/std': 0.21246552467346191, 'reward': 0.23267492651939392, 'reward_std': 0.21246552467346191, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20715653896331787, 'sampling/sampling_logp_difference/max': 1.215444564819336, 'sampling/importance_sampling_ratio/min': 0.29657813906669617, 'sampling/importance_sampling_ratio/mean': 1.0435240268707275, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.596053183078766, 'clip_ratio/low_mean': 0.07777901645749807, 'clip_ratio/low_min': 0.07777901645749807, 'clip_ratio/high_mean': 0.1069975420832634, 'clip_ratio/high_max': 0.1069975420832634, 'clip_ratio/region_mean': 0.18477655854076147, 'reward_total_mean': 0.23267492651939392, 'reward_meter_mean': 0.48370596766471863, 'reward_meter_std': 0.41808390617370605, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9972826242446899, 'reward_lexical_plausibility_std': 0.007685941644012928, 'reward_repeat_penalty_mean': 0.9845765829086304, 'reward_repeat_penalty_std': 0.01943829655647278, 'reward_near_duplicate_penalty_mean': 0.9938940405845642, 'reward_near_duplicate_penalty_std': 0.003942488227039576, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9905967712402344, 'reward_distinct_2_std': 0.017439009621739388, 'reward_judge_quality_mean': 0.44999998807907104, 'reward_judge_quality_std': 0.09258200973272324, 'reward_total_composite_mean': 0.23267492651939392, 'reward_total_composite_std': 0.21246552467346191, 'epoch': 0.01}

  6%|▌         | 188/3300 [39:42<8:29:28,  9.82s/it]INFO 04-12 11:25:54 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 189/3300 [39:54<9:08:31, 10.58s/it]2026-04-12 11:26:06,524 | INFO | train_grpo_train | metrics_logged mode=train step=189

                                                    
{'loss': 0.024, 'grad_norm': 13.126042366027832, 'learning_rate': 9.43030303030303e-06, 'num_tokens': 398385.0, 'completions/mean_length': 58.25, 'completions/min_length': 53.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.25, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.5347265005111694, 'rewards/meter/std': 0.34641000628471375, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.47874999046325684, 'rewards/judge_quality/std': 0.1683056354522705, 'rewards/repeat_penalty/mean': 0.9957600831985474, 'rewards/repeat_penalty/std': 0.0037746031302958727, 'rewards/near_duplicate_penalty/mean': 0.9957600831985474, 'rewards/near_duplicate_penalty/std': 0.0037746031302958727, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.20895373821258545, 'rewards/total_composite/std': 0.17416755855083466, 'reward': 0.20895373821258545, 'reward_std': 0.17416755855083466, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1976327896118164, 'sampling/sampling_logp_difference/max': 1.7564430236816406, 'sampling/importance_sampling_ratio/min': 0.17265790700912476, 'sampling/importance_sampling_ratio/mean': 1.0394078493118286, 'sampling/importance_sampling_ratio/max': 1.98775053024292, 'entropy': 2.3172446340322495, 'clip_ratio/low_mean': 0.08750543650239706, 'clip_ratio/low_min': 0.08750543650239706, 'clip_ratio/high_mean': 0.0921156071126461, 'clip_ratio/high_max': 0.0921156071126461, 'clip_ratio/region_mean': 0.17962104361504316, 'reward_total_mean': 0.20895373821258545, 'reward_meter_mean': 0.5347265005111694, 'reward_meter_std': 0.34641000628471375, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9957600831985474, 'reward_repeat_penalty_std': 0.0037746031302958727, 'reward_near_duplicate_penalty_mean': 0.9957600831985474, 'reward_near_duplicate_penalty_std': 0.0037746031302958727, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.47874999046325684, 'reward_judge_quality_std': 0.1683056354522705, 'reward_total_composite_mean': 0.20895373821258545, 'reward_total_composite_std': 0.17416755855083466, 'epoch': 0.01}

  6%|▌         | 189/3300 [39:54<9:08:31, 10.58s/it]INFO 04-12 11:26:06 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 190/3300 [40:07<9:43:44, 11.26s/it]2026-04-12 11:26:19,378 | INFO | train_grpo_train | metrics_logged mode=train step=190

                                                    
{'loss': 0.0626, 'grad_norm': 8.389774322509766, 'learning_rate': 9.427272727272728e-06, 'num_tokens': 401256.0, 'completions/mean_length': 140.875, 'completions/min_length': 112.0, 'completions/max_length': 181.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 140.875, 'completions/min_terminated_length': 112.0, 'completions/max_terminated_length': 181.0, 'rewards/meter/mean': 0.26289302110671997, 'rewards/meter/std': 0.16595962643623352, 'rewards/count_adherence/mean': 0.910714328289032, 'rewards/count_adherence/std': 0.07393559068441391, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9588995575904846, 'rewards/lexical_plausibility/std': 0.06422513723373413, 'rewards/judge_quality/mean': 0.2875000238418579, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.993451714515686, 'rewards/repeat_penalty/std': 0.013639457523822784, 'rewards/near_duplicate_penalty/mean': 0.9976036548614502, 'rewards/near_duplicate_penalty/std': 0.0023936352226883173, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9958193898200989, 'rewards/distinct_2/std': 0.0118245305493474, 'rewards/total_composite/mean': 0.06927891075611115, 'rewards/total_composite/std': 0.06271076202392578, 'reward': 0.06927891075611115, 'reward_std': 0.06271075457334518, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2442931979894638, 'sampling/sampling_logp_difference/max': 1.2616536617279053, 'sampling/importance_sampling_ratio/min': 0.28318536281585693, 'sampling/importance_sampling_ratio/mean': 1.0808517932891846, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.8979704678058624, 'clip_ratio/low_mean': 0.12465440481901169, 'clip_ratio/low_min': 0.12465440481901169, 'clip_ratio/high_mean': 0.07201832346618176, 'clip_ratio/high_max': 0.07201832346618176, 'clip_ratio/region_mean': 0.19667272828519344, 'reward_total_mean': 0.06927891075611115, 'reward_meter_mean': 0.26289302110671997, 'reward_meter_std': 0.16595962643623352, 'reward_count_adherence_mean': 0.910714328289032, 'reward_count_adherence_std': 0.07393559068441391, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9588995575904846, 'reward_lexical_plausibility_std': 0.06422513723373413, 'reward_repeat_penalty_mean': 0.993451714515686, 'reward_repeat_penalty_std': 0.013639457523822784, 'reward_near_duplicate_penalty_mean': 0.9976036548614502, 'reward_near_duplicate_penalty_std': 0.0023936352226883173, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9958193898200989, 'reward_distinct_2_std': 0.0118245305493474, 'reward_judge_quality_mean': 0.2875000238418579, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.06927891075611115, 'reward_total_composite_std': 0.06271076202392578, 'epoch': 0.01}

  6%|▌         | 190/3300 [40:07<9:43:44, 11.26s/it]INFO 04-12 11:26:19 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 191/3300 [40:19<9:53:50, 11.46s/it]2026-04-12 11:26:31,302 | INFO | train_grpo_train | metrics_logged mode=train step=191

                                                    
{'loss': 0.073, 'grad_norm': 11.953428268432617, 'learning_rate': 9.424242424242425e-06, 'num_tokens': 403246.0, 'completions/mean_length': 67.75, 'completions/min_length': 63.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.75, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.4731566905975342, 'rewards/meter/std': 0.4082699120044708, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4374999701976776, 'rewards/judge_quality/std': 0.1642080396413803, 'rewards/repeat_penalty/mean': 0.9625617265701294, 'rewards/repeat_penalty/std': 0.04136313870549202, 'rewards/near_duplicate_penalty/mean': 0.9850481748580933, 'rewards/near_duplicate_penalty/std': 0.01034289225935936, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9771062135696411, 'rewards/distinct_2/std': 0.0388520210981369, 'rewards/total_composite/mean': 0.2468075454235077, 'rewards/total_composite/std': 0.26338961720466614, 'reward': 0.2468075454235077, 'reward_std': 0.26338958740234375, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23241984844207764, 'sampling/sampling_logp_difference/max': 1.9645910263061523, 'sampling/importance_sampling_ratio/min': 0.1402132213115692, 'sampling/importance_sampling_ratio/mean': 1.0427559614181519, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.699978247284889, 'clip_ratio/low_mean': 0.10380714759230614, 'clip_ratio/low_min': 0.10380714759230614, 'clip_ratio/high_mean': 0.07659748196601868, 'clip_ratio/high_max': 0.07659748196601868, 'clip_ratio/region_mean': 0.1804046295583248, 'reward_total_mean': 0.2468075454235077, 'reward_meter_mean': 0.4731566905975342, 'reward_meter_std': 0.4082699120044708, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9625617265701294, 'reward_repeat_penalty_std': 0.04136313870549202, 'reward_near_duplicate_penalty_mean': 0.9850481748580933, 'reward_near_duplicate_penalty_std': 0.01034289225935936, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9771062135696411, 'reward_distinct_2_std': 0.0388520210981369, 'reward_judge_quality_mean': 0.4374999701976776, 'reward_judge_quality_std': 0.1642080396413803, 'reward_total_composite_mean': 0.2468075454235077, 'reward_total_composite_std': 0.26338961720466614, 'epoch': 0.01}

  6%|▌         | 191/3300 [40:19<9:53:50, 11.46s/it]INFO 04-12 11:26:31 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 192/3300 [40:33<10:39:16, 12.34s/it]2026-04-12 11:26:45,698 | INFO | train_grpo_train | metrics_logged mode=train step=192

                                                     
{'loss': 0.0741, 'grad_norm': 8.319887161254883, 'learning_rate': 9.421212121212122e-06, 'num_tokens': 405928.0, 'completions/mean_length': 149.25, 'completions/min_length': 117.0, 'completions/max_length': 186.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 149.25, 'completions/min_terminated_length': 117.0, 'completions/max_terminated_length': 186.0, 'rewards/meter/mean': 0.6476424336433411, 'rewards/meter/std': 0.2817014753818512, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.9902408123016357, 'rewards/repeat_penalty/std': 0.011109532788395882, 'rewards/near_duplicate_penalty/mean': 0.9966154098510742, 'rewards/near_duplicate_penalty/std': 0.003634650958701968, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9935892224311829, 'rewards/distinct_2/std': 0.008848296478390694, 'rewards/total_composite/mean': 0.18223273754119873, 'rewards/total_composite/std': 0.11901077628135681, 'reward': 0.18223273754119873, 'reward_std': 0.11901078373193741, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2439643144607544, 'sampling/sampling_logp_difference/max': 1.388969898223877, 'sampling/importance_sampling_ratio/min': 0.2493320107460022, 'sampling/importance_sampling_ratio/mean': 1.0756523609161377, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 4.0335632264614105, 'clip_ratio/low_mean': 0.09023445472121239, 'clip_ratio/low_min': 0.09023445472121239, 'clip_ratio/high_mean': 0.12000403553247452, 'clip_ratio/high_max': 0.12000403553247452, 'clip_ratio/region_mean': 0.2102384902536869, 'reward_total_mean': 0.18223273754119873, 'reward_meter_mean': 0.6476424336433411, 'reward_meter_std': 0.2817014753818512, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9902408123016357, 'reward_repeat_penalty_std': 0.011109532788395882, 'reward_near_duplicate_penalty_mean': 0.9966154098510742, 'reward_near_duplicate_penalty_std': 0.003634650958701968, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9935892224311829, 'reward_distinct_2_std': 0.008848296478390694, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.18223273754119873, 'reward_total_composite_std': 0.11901077628135681, 'epoch': 0.01}

  6%|▌         | 192/3300 [40:33<10:39:16, 12.34s/it]INFO 04-12 11:26:45 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 193/3300 [40:44<10:10:59, 11.80s/it]2026-04-12 11:26:56,233 | INFO | train_grpo_train | metrics_logged mode=train step=193

                                                     
{'loss': -0.0772, 'grad_norm': 16.670808792114258, 'learning_rate': 9.418181818181818e-06, 'num_tokens': 407423.0, 'completions/mean_length': 39.875, 'completions/min_length': 32.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.875, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.38676023483276367, 'rewards/meter/std': 0.40895742177963257, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3712500035762787, 'rewards/judge_quality/std': 0.25670650601387024, 'rewards/repeat_penalty/mean': 0.9810949563980103, 'rewards/repeat_penalty/std': 0.046554312109947205, 'rewards/near_duplicate_penalty/mean': 0.99781733751297, 'rewards/near_duplicate_penalty/std': 0.00369036546908319, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9832775592803955, 'rewards/distinct_2/std': 0.0472981221973896, 'rewards/total_composite/mean': 0.15096087753772736, 'rewards/total_composite/std': 0.1640724241733551, 'reward': 0.15096087753772736, 'reward_std': 0.1640724241733551, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.25328776240348816, 'sampling/sampling_logp_difference/max': 1.7909431457519531, 'sampling/importance_sampling_ratio/min': 0.166802778840065, 'sampling/importance_sampling_ratio/mean': 1.0317927598953247, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.5228359401226044, 'clip_ratio/low_mean': 0.12745822221040726, 'clip_ratio/low_min': 0.12745822221040726, 'clip_ratio/high_mean': 0.0722408052533865, 'clip_ratio/high_max': 0.0722408052533865, 'clip_ratio/region_mean': 0.19969902746379375, 'reward_total_mean': 0.15096087753772736, 'reward_meter_mean': 0.38676023483276367, 'reward_meter_std': 0.40895742177963257, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9810949563980103, 'reward_repeat_penalty_std': 0.046554312109947205, 'reward_near_duplicate_penalty_mean': 0.99781733751297, 'reward_near_duplicate_penalty_std': 0.00369036546908319, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9832775592803955, 'reward_distinct_2_std': 0.0472981221973896, 'reward_judge_quality_mean': 0.3712500035762787, 'reward_judge_quality_std': 0.25670650601387024, 'reward_total_composite_mean': 0.15096087753772736, 'reward_total_composite_std': 0.1640724241733551, 'epoch': 0.01}

  6%|▌         | 193/3300 [40:44<10:10:59, 11.80s/it]INFO 04-12 11:26:56 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 194/3300 [40:55<10:03:50, 11.66s/it]2026-04-12 11:27:07,583 | INFO | train_grpo_train | metrics_logged mode=train step=194

                                                     
{'loss': 0.0542, 'grad_norm': 7.748849868774414, 'learning_rate': 9.415151515151515e-06, 'num_tokens': 410166.0, 'completions/mean_length': 148.875, 'completions/min_length': 129.0, 'completions/max_length': 177.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 148.875, 'completions/min_terminated_length': 129.0, 'completions/max_terminated_length': 177.0, 'rewards/meter/mean': 0.46342623233795166, 'rewards/meter/std': 0.29466667771339417, 'rewards/count_adherence/mean': 0.8928571939468384, 'rewards/count_adherence/std': 0.10101525485515594, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.9451513886451721, 'rewards/lexical_plausibility/std': 0.08803725242614746, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.993679404258728, 'rewards/repeat_penalty/std': 0.008588011376559734, 'rewards/near_duplicate_penalty/mean': 0.996414065361023, 'rewards/near_duplicate_penalty/std': 0.002788442187011242, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9972527623176575, 'rewards/distinct_2/std': 0.007770403753966093, 'rewards/total_composite/mean': 0.09386736899614334, 'rewards/total_composite/std': 0.1131976842880249, 'reward': 0.09386736899614334, 'reward_std': 0.1131976842880249, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.25746795535087585, 'sampling/sampling_logp_difference/max': 2.3538103103637695, 'sampling/importance_sampling_ratio/min': 0.09500646591186523, 'sampling/importance_sampling_ratio/mean': 1.0607352256774902, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.6030998826026917, 'clip_ratio/low_mean': 0.16044802591204643, 'clip_ratio/low_min': 0.16044802591204643, 'clip_ratio/high_mean': 0.05099865794181824, 'clip_ratio/high_max': 0.05099865794181824, 'clip_ratio/region_mean': 0.21144668385386467, 'reward_total_mean': 0.09386736899614334, 'reward_meter_mean': 0.46342623233795166, 'reward_meter_std': 0.29466667771339417, 'reward_count_adherence_mean': 0.8928571939468384, 'reward_count_adherence_std': 0.10101525485515594, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.9451513886451721, 'reward_lexical_plausibility_std': 0.08803725242614746, 'reward_repeat_penalty_mean': 0.993679404258728, 'reward_repeat_penalty_std': 0.008588011376559734, 'reward_near_duplicate_penalty_mean': 0.996414065361023, 'reward_near_duplicate_penalty_std': 0.002788442187011242, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9972527623176575, 'reward_distinct_2_std': 0.007770403753966093, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.09386736899614334, 'reward_total_composite_std': 0.1131976842880249, 'epoch': 0.01}

  6%|▌         | 194/3300 [40:55<10:03:50, 11.66s/it]INFO 04-12 11:27:07 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 195/3300 [41:05<9:41:34, 11.24s/it] 2026-04-12 11:27:17,825 | INFO | train_grpo_train | metrics_logged mode=train step=195

                                                    
{'loss': 0.0599, 'grad_norm': 11.657549858093262, 'learning_rate': 9.412121212121212e-06, 'num_tokens': 412182.0, 'completions/mean_length': 76.0, 'completions/min_length': 68.0, 'completions/max_length': 83.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.0, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.4964779317378998, 'rewards/meter/std': 0.23041903972625732, '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/lexical_plausibility/mean': 0.9986978769302368, 'rewards/lexical_plausibility/std': 0.0036828548181802034, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.05345224589109421, 'rewards/repeat_penalty/mean': 0.9933637976646423, 'rewards/repeat_penalty/std': 0.005556311458349228, 'rewards/near_duplicate_penalty/mean': 0.9933637976646423, 'rewards/near_duplicate_penalty/std': 0.005556311458349228, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.20080244541168213, 'rewards/total_composite/std': 0.10564137250185013, 'reward': 0.20080244541168213, 'reward_std': 0.10564136505126953, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22843697667121887, 'sampling/sampling_logp_difference/max': 1.997262954711914, 'sampling/importance_sampling_ratio/min': 0.13570620119571686, 'sampling/importance_sampling_ratio/mean': 1.049808144569397, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.66208553314209, 'clip_ratio/low_mean': 0.09703397192060947, 'clip_ratio/low_min': 0.09703397192060947, 'clip_ratio/high_mean': 0.08621727302670479, 'clip_ratio/high_max': 0.08621727302670479, 'clip_ratio/region_mean': 0.18325124494731426, 'reward_total_mean': 0.20080244541168213, 'reward_meter_mean': 0.4964779317378998, 'reward_meter_std': 0.23041903972625732, '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_lexical_plausibility_mean': 0.9986978769302368, 'reward_lexical_plausibility_std': 0.0036828548181802034, 'reward_repeat_penalty_mean': 0.9933637976646423, 'reward_repeat_penalty_std': 0.005556311458349228, 'reward_near_duplicate_penalty_mean': 0.9933637976646423, 'reward_near_duplicate_penalty_std': 0.005556311458349228, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.05345224589109421, 'reward_total_composite_mean': 0.20080244541168213, 'reward_total_composite_std': 0.10564137250185013, 'epoch': 0.01}

  6%|▌         | 195/3300 [41:05<9:41:34, 11.24s/it]INFO 04-12 11:27:18 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 196/3300 [41:14<9:09:32, 10.62s/it]2026-04-12 11:27:27,013 | INFO | train_grpo_train | metrics_logged mode=train step=196

                                                    
{'loss': 0.0102, 'grad_norm': 15.35546588897705, 'learning_rate': 9.40909090909091e-06, 'num_tokens': 413508.0, 'completions/mean_length': 15.75, 'completions/min_length': 14.0, 'completions/max_length': 22.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 15.75, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 22.0, 'rewards/meter/mean': 0.9720182418823242, 'rewards/meter/std': 0.018872657790780067, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7775000333786011, 'rewards/judge_quality/std': 0.26385873556137085, 'rewards/repeat_penalty/mean': 0.6641231775283813, 'rewards/repeat_penalty/std': 0.09269300103187561, 'rewards/near_duplicate_penalty/mean': 0.8854975700378418, 'rewards/near_duplicate_penalty/std': 0.12359067052602768, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.7585351467132568, 'rewards/total_composite/std': 0.26263073086738586, 'reward': 0.7585351467132568, 'reward_std': 0.26263073086738586, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15669620037078857, 'sampling/sampling_logp_difference/max': 1.0087833404541016, 'sampling/importance_sampling_ratio/min': 0.36466237902641296, 'sampling/importance_sampling_ratio/mean': 1.0573863983154297, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0078484639525414, 'clip_ratio/low_mean': 0.033333334140479565, 'clip_ratio/low_min': 0.033333334140479565, 'clip_ratio/high_mean': 0.07552760001271963, 'clip_ratio/high_max': 0.07552760001271963, 'clip_ratio/region_mean': 0.1088609341531992, 'reward_total_mean': 0.7585351467132568, 'reward_meter_mean': 0.9720182418823242, 'reward_meter_std': 0.018872657790780067, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6641231775283813, 'reward_repeat_penalty_std': 0.09269300103187561, 'reward_near_duplicate_penalty_mean': 0.8854975700378418, 'reward_near_duplicate_penalty_std': 0.12359067052602768, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7775000333786011, 'reward_judge_quality_std': 0.26385873556137085, 'reward_total_composite_mean': 0.7585351467132568, 'reward_total_composite_std': 0.26263073086738586, 'epoch': 0.01}

  6%|▌         | 196/3300 [41:14<9:09:32, 10.62s/it]INFO 04-12 11:27:27 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 197/3300 [41:32<10:54:19, 12.65s/it]2026-04-12 11:27:44,401 | INFO | train_grpo_train | metrics_logged mode=train step=197

                                                     
{'loss': -0.0622, 'grad_norm': 7.669545650482178, 'learning_rate': 9.406060606060607e-06, 'num_tokens': 416069.0, 'completions/mean_length': 197.125, 'completions/min_length': 122.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 152.1428680419922, 'completions/min_terminated_length': 122.0, 'completions/max_terminated_length': 173.0, 'rewards/meter/mean': 0.2418033331632614, 'rewards/meter/std': 0.10072522610425949, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.06681530922651291, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/lexical_plausibility/mean': 0.9564472436904907, 'rewards/lexical_plausibility/std': 0.11238300800323486, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.9782454371452332, 'rewards/repeat_penalty/std': 0.025253357365727425, 'rewards/near_duplicate_penalty/mean': 0.9932018518447876, 'rewards/near_duplicate_penalty/std': 0.0027052685618400574, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.984892725944519, 'rewards/distinct_2/std': 0.023083476349711418, 'rewards/total_composite/mean': 0.04156244173645973, 'rewards/total_composite/std': 0.057584863156080246, 'reward': 0.04156244173645973, 'reward_std': 0.05758485570549965, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23801890015602112, 'sampling/sampling_logp_difference/max': 1.5554094314575195, 'sampling/importance_sampling_ratio/min': 0.2111029326915741, 'sampling/importance_sampling_ratio/mean': 1.0642399787902832, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.8230305314064026, 'clip_ratio/low_mean': 0.10472686402499676, 'clip_ratio/low_min': 0.10472686402499676, 'clip_ratio/high_mean': 0.0750449001789093, 'clip_ratio/high_max': 0.0750449001789093, 'clip_ratio/region_mean': 0.17977176420390606, 'reward_total_mean': 0.04156244173645973, 'reward_meter_mean': 0.2418033331632614, 'reward_meter_std': 0.10072522610425949, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.06681530922651291, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_lexical_plausibility_mean': 0.9564472436904907, 'reward_lexical_plausibility_std': 0.11238300800323486, 'reward_repeat_penalty_mean': 0.9782454371452332, 'reward_repeat_penalty_std': 0.025253357365727425, 'reward_near_duplicate_penalty_mean': 0.9932018518447876, 'reward_near_duplicate_penalty_std': 0.0027052685618400574, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.984892725944519, 'reward_distinct_2_std': 0.023083476349711418, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.04156244173645973, 'reward_total_composite_std': 0.057584863156080246, 'epoch': 0.01}

  6%|▌         | 197/3300 [41:32<10:54:19, 12.65s/it]INFO 04-12 11:27:44 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 198/3300 [41:43<10:29:39, 12.18s/it]2026-04-12 11:27:55,480 | INFO | train_grpo_train | metrics_logged mode=train step=198

                                                     
{'loss': 0.075, 'grad_norm': 15.811528205871582, 'learning_rate': 9.403030303030304e-06, 'num_tokens': 417835.0, 'completions/mean_length': 52.75, 'completions/min_length': 45.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.75, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.44315749406814575, 'rewards/meter/std': 0.33838579058647156, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9977678656578064, 'rewards/lexical_plausibility/std': 0.006313450634479523, 'rewards/judge_quality/mean': 0.4662500023841858, 'rewards/judge_quality/std': 0.2180391401052475, 'rewards/repeat_penalty/mean': 0.9951223134994507, 'rewards/repeat_penalty/std': 0.010539373382925987, 'rewards/near_duplicate_penalty/mean': 0.9951223134994507, 'rewards/near_duplicate_penalty/std': 0.010539373382925987, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.16395926475524902, 'rewards/total_composite/std': 0.1340993195772171, 'reward': 0.16395926475524902, 'reward_std': 0.1340993195772171, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2545931339263916, 'sampling/sampling_logp_difference/max': 2.2817153930664062, 'sampling/importance_sampling_ratio/min': 0.102108895778656, 'sampling/importance_sampling_ratio/mean': 1.0445752143859863, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.5573186725378036, 'clip_ratio/low_mean': 0.10163934528827667, 'clip_ratio/low_min': 0.10163934528827667, 'clip_ratio/high_mean': 0.08619281370192766, 'clip_ratio/high_max': 0.08619281370192766, 'clip_ratio/region_mean': 0.18783215899020433, 'reward_total_mean': 0.16395926475524902, 'reward_meter_mean': 0.44315749406814575, 'reward_meter_std': 0.33838579058647156, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9977678656578064, 'reward_lexical_plausibility_std': 0.006313450634479523, 'reward_repeat_penalty_mean': 0.9951223134994507, 'reward_repeat_penalty_std': 0.010539373382925987, 'reward_near_duplicate_penalty_mean': 0.9951223134994507, 'reward_near_duplicate_penalty_std': 0.010539373382925987, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4662500023841858, 'reward_judge_quality_std': 0.2180391401052475, 'reward_total_composite_mean': 0.16395926475524902, 'reward_total_composite_std': 0.1340993195772171, 'epoch': 0.01}

  6%|▌         | 198/3300 [41:43<10:29:39, 12.18s/it]INFO 04-12 11:27:55 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 199/3300 [41:53<10:00:37, 11.62s/it]2026-04-12 11:28:05,796 | INFO | train_grpo_train | metrics_logged mode=train step=199

                                                     
{'loss': 0.0383, 'grad_norm': 17.07819175720215, 'learning_rate': 9.4e-06, 'num_tokens': 419442.0, 'completions/mean_length': 37.875, 'completions/min_length': 35.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.875, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.8455114364624023, 'rewards/meter/std': 0.33083054423332214, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9853488206863403, 'rewards/lexical_plausibility/std': 0.034790314733982086, 'rewards/judge_quality/mean': 0.4412499964237213, 'rewards/judge_quality/std': 0.1829470992088318, 'rewards/repeat_penalty/mean': 0.9876461029052734, 'rewards/repeat_penalty/std': 0.018724631518125534, 'rewards/near_duplicate_penalty/mean': 0.9876461029052734, 'rewards/near_duplicate_penalty/std': 0.018724631518125534, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3867346942424774, 'rewards/total_composite/std': 0.23983639478683472, 'reward': 0.3867346942424774, 'reward_std': 0.23983637988567352, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22389063239097595, 'sampling/sampling_logp_difference/max': 1.426994800567627, 'sampling/importance_sampling_ratio/min': 0.24002918601036072, 'sampling/importance_sampling_ratio/mean': 1.045613408088684, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.696754664182663, 'clip_ratio/low_mean': 0.10653246194124222, 'clip_ratio/low_min': 0.10653246194124222, 'clip_ratio/high_mean': 0.06676681339740753, 'clip_ratio/high_max': 0.06676681339740753, 'clip_ratio/region_mean': 0.17329927533864975, 'reward_total_mean': 0.3867346942424774, 'reward_meter_mean': 0.8455114364624023, 'reward_meter_std': 0.33083054423332214, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9853488206863403, 'reward_lexical_plausibility_std': 0.034790314733982086, 'reward_repeat_penalty_mean': 0.9876461029052734, 'reward_repeat_penalty_std': 0.018724631518125534, 'reward_near_duplicate_penalty_mean': 0.9876461029052734, 'reward_near_duplicate_penalty_std': 0.018724631518125534, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4412499964237213, 'reward_judge_quality_std': 0.1829470992088318, 'reward_total_composite_mean': 0.3867346942424774, 'reward_total_composite_std': 0.23983639478683472, 'epoch': 0.01}

  6%|▌         | 199/3300 [41:53<10:00:37, 11.62s/it]INFO 04-12 11:28:06 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 200/3300 [42:02<9:12:40, 10.70s/it] 2026-04-12 11:28:14,336 | INFO | train_grpo_train | metrics_logged mode=train step=200

                                                    
{'loss': -0.002, 'grad_norm': 12.940689086914062, 'learning_rate': 9.396969696969697e-06, 'num_tokens': 420952.0, 'completions/mean_length': 27.75, 'completions/min_length': 26.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.75, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.8510260581970215, 'rewards/meter/std': 0.3145773410797119, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.9624999761581421, 'rewards/judge_quality/std': 0.023145509883761406, 'rewards/repeat_penalty/mean': 0.9975649118423462, 'rewards/repeat_penalty/std': 0.006887406576424837, 'rewards/near_duplicate_penalty/mean': 0.9975649118423462, 'rewards/near_duplicate_penalty/std': 0.006887406576424837, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.820294201374054, 'rewards/total_composite/std': 0.3036881983280182, 'reward': 0.820294201374054, 'reward_std': 0.3036881983280182, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06326823681592941, 'sampling/sampling_logp_difference/max': 0.9554257392883301, 'sampling/importance_sampling_ratio/min': 0.3846483826637268, 'sampling/importance_sampling_ratio/mean': 1.0123517513275146, 'sampling/importance_sampling_ratio/max': 1.8747886419296265, 'entropy': 0.29799419827759266, 'clip_ratio/low_mean': 0.008928571827709675, 'clip_ratio/low_min': 0.008928571827709675, 'clip_ratio/high_mean': 0.04856243031099439, 'clip_ratio/high_max': 0.04856243031099439, 'clip_ratio/region_mean': 0.05749100213870406, 'reward_total_mean': 0.820294201374054, 'reward_meter_mean': 0.8510260581970215, 'reward_meter_std': 0.3145773410797119, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9975649118423462, 'reward_repeat_penalty_std': 0.006887406576424837, 'reward_near_duplicate_penalty_mean': 0.9975649118423462, 'reward_near_duplicate_penalty_std': 0.006887406576424837, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.9624999761581421, 'reward_judge_quality_std': 0.023145509883761406, 'reward_total_composite_mean': 0.820294201374054, 'reward_total_composite_std': 0.3036881983280182, 'epoch': 0.01}

  6%|▌         | 200/3300 [42:02<9:12:40, 10.70s/it]INFO 04-12 11:28:14 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


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

 15%|█▌        | 2/13 [00:11<01:03,  5.82s/it]

 23%|██▎       | 3/13 [00:19<01:05,  6.51s/it]

 31%|███       | 4/13 [00:29<01:10,  7.82s/it]

 38%|███▊      | 5/13 [00:38<01:06,  8.29s/it]

 46%|████▌     | 6/13 [00:47<01:00,  8.67s/it]

 54%|█████▍    | 7/13 [00:56<00:52,  8.68s/it]

 62%|██████▏   | 8/13 [01:06<00:45,  9.12s/it]

 69%|██████▉   | 9/13 [01:18<00:39,  9.90s/it]

 77%|███████▋  | 10/13 [01:28<00:29,  9.94s/it]

 85%|████████▍ | 11/13 [01:37<00:19,  9.76s/it]

 92%|█████████▏| 12/13 [01:50<00:10, 10.67s/it]

100%|██████████| 13/13 [02:02<00:00, 11.22s/it]2026-04-12 11:30:29,481 | INFO | train_grpo_train | metrics_logged mode=eval step=200

                                                    

                                               
{'eval_loss': nan, 'eval_runtime': 135.1266, 'eval_samples_per_second': 0.77, 'eval_steps_per_second': 0.096, 'eval_num_tokens': 420952.0, 'eval_completions/mean_length': 136.65384615384616, 'eval_completions/min_length': 34.07692307692308, 'eval_completions/max_length': 326.2307692307692, 'eval_completions/clipped_ratio': 0.009615384615384616, 'eval_completions/mean_terminated_length': 132.82005544809195, 'eval_completions/min_terminated_length': 34.07692307692308, 'eval_completions/max_terminated_length': 317.15384615384613, 'eval_rewards/meter/mean': 0.30902680640037244, 'eval_rewards/meter/std': 0.3418757256407004, 'eval_rewards/count_adherence/mean': 0.9369382766576914, 'eval_rewards/count_adherence/std': 0.1166658506083947, 'eval_rewards/arabic_clean/mean': 0.8461538461538461, 'eval_rewards/arabic_clean/std': 0.314018373305981, 'eval_rewards/lexical_plausibility/mean': 0.9651182614839994, 'eval_rewards/lexical_plausibility/std': 0.07399109987398753, 'eval_rewards/judge_quality/mean': 0.3369230765562791, 'eval_rewards/judge_quality/std': 0.1494664383622316, 'eval_rewards/repeat_penalty/mean': 0.9602930866754972, 'eval_rewards/repeat_penalty/std': 0.06996360016413607, 'eval_rewards/near_duplicate_penalty/mean': 0.9855904900110685, 'eval_rewards/near_duplicate_penalty/std': 0.025130621175496623, 'eval_rewards/opening_diversity/mean': 0.9879807692307693, 'eval_rewards/opening_diversity/std': 0.03399551774446781, 'eval_rewards/distinct_2/mean': 0.9864450830679673, 'eval_rewards/distinct_2/std': 0.02632404539662485, 'eval_rewards/total_composite/mean': 0.09235788251344974, 'eval_rewards/total_composite/std': 0.12527622626377985, 'eval_reward': 0.09235788251344974, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.17638242817842043, 'eval_sampling/sampling_logp_difference/max': 1.21868713085468, 'eval_sampling/importance_sampling_ratio/min': 0.2973355467502887, 'eval_sampling/importance_sampling_ratio/mean': 1.058694784457867, 'eval_sampling/importance_sampling_ratio/max': 1.7514545642412627, 'eval_entropy': 3.374564317556528, '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': 0.09235788251344974, 'eval_reward_meter_mean': 0.30902680640037244, 'eval_reward_meter_std': 0.3418757256407004, 'eval_reward_count_adherence_mean': 0.9369382766576914, 'eval_reward_count_adherence_std': 0.1166658506083947, 'eval_reward_arabic_clean_mean': 0.8461538461538461, 'eval_reward_arabic_clean_std': 0.314018373305981, 'eval_reward_lexical_plausibility_mean': 0.9651182614839994, 'eval_reward_lexical_plausibility_std': 0.07399109987398753, 'eval_reward_repeat_penalty_mean': 0.9602930866754972, 'eval_reward_repeat_penalty_std': 0.06996360016413607, 'eval_reward_near_duplicate_penalty_mean': 0.9855904900110685, 'eval_reward_near_duplicate_penalty_std': 0.025130621175496623, 'eval_reward_opening_diversity_mean': 0.9879807692307693, 'eval_reward_opening_diversity_std': 0.03399551774446781, 'eval_reward_distinct_2_mean': 0.9864450830679673, 'eval_reward_distinct_2_std': 0.02632404539662485, 'eval_reward_judge_quality_mean': 0.3369230765562791, 'eval_reward_judge_quality_std': 0.1494664383622316, 'eval_reward_total_composite_mean': 0.09235788251344974, 'eval_reward_total_composite_std': 0.12527622626377985, 'epoch': 0.01}

  6%|▌         | 200/3300 [44:17<9:12:40, 10.70s/it]

100%|██████████| 13/13 [02:02<00:00, 11.22s/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(

  6%|▌         | 201/3300 [44:29<44:22:54, 51.56s/it]2026-04-12 11:30:41,239 | INFO | train_grpo_train | metrics_logged mode=train step=201

                                                     
{'loss': 0.0372, 'grad_norm': 25.896852493286133, 'learning_rate': 9.393939393939396e-06, 'num_tokens': 422224.0, 'completions/mean_length': 15.0, 'completions/min_length': 13.0, 'completions/max_length': 17.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 15.0, 'completions/min_terminated_length': 13.0, 'completions/max_terminated_length': 17.0, 'rewards/meter/mean': 0.12249067425727844, 'rewards/meter/std': 0.34225091338157654, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.5345224738121033, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.8237500190734863, 'rewards/judge_quality/std': 0.2722361385822296, 'rewards/repeat_penalty/mean': 0.875, 'rewards/repeat_penalty/std': 0.13363061845302582, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.875, 'rewards/opening_diversity/std': 0.13363061845302582, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.11268244683742523, 'rewards/total_composite/std': 0.3148745000362396, 'reward': 0.11268244683742523, 'reward_std': 0.3148745000362396, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.266227126121521, 'sampling/sampling_logp_difference/max': 1.3496150970458984, 'sampling/importance_sampling_ratio/min': 0.2593400478363037, 'sampling/importance_sampling_ratio/mean': 1.0251189470291138, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0586407631635666, 'clip_ratio/low_mean': 0.23556076735258102, 'clip_ratio/low_min': 0.23556076735258102, 'clip_ratio/high_mean': 0.03333333507180214, 'clip_ratio/high_max': 0.03333333507180214, 'clip_ratio/region_mean': 0.26889410242438316, 'reward_total_mean': 0.11268244683742523, 'reward_meter_mean': 0.12249067425727844, 'reward_meter_std': 0.34225091338157654, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.5345224738121033, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.875, 'reward_repeat_penalty_std': 0.13363061845302582, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.875, 'reward_opening_diversity_std': 0.13363061845302582, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8237500190734863, 'reward_judge_quality_std': 0.2722361385822296, 'reward_total_composite_mean': 0.11268244683742523, 'reward_total_composite_std': 0.3148745000362396, 'epoch': 0.01}

  6%|▌         | 201/3300 [44:29<44:22:54, 51.56s/it]INFO 04-12 11:30:41 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 202/3300 [44:40<34:03:20, 39.57s/it]2026-04-12 11:30:52,848 | INFO | train_grpo_train | metrics_logged mode=train step=202

                                                     
{'loss': 0.0907, 'grad_norm': 19.980554580688477, 'learning_rate': 9.390909090909092e-06, 'num_tokens': 424013.0, 'completions/mean_length': 58.625, 'completions/min_length': 41.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.625, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.48397544026374817, 'rewards/meter/std': 0.2964073717594147, '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/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.48374998569488525, 'rewards/judge_quality/std': 0.18196842074394226, 'rewards/repeat_penalty/mean': 0.9771234393119812, 'rewards/repeat_penalty/std': 0.026555240154266357, 'rewards/near_duplicate_penalty/mean': 0.9825642704963684, 'rewards/near_duplicate_penalty/std': 0.014555198140442371, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9943438768386841, 'rewards/distinct_2/std': 0.015997890383005142, 'rewards/total_composite/mean': 0.20144273340702057, 'rewards/total_composite/std': 0.13822034001350403, 'reward': 0.20144273340702057, 'reward_std': 0.13822034001350403, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.26058104634284973, 'sampling/sampling_logp_difference/max': 2.380445957183838, 'sampling/importance_sampling_ratio/min': 0.09250931441783905, 'sampling/importance_sampling_ratio/mean': 0.9931546449661255, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1785653233528137, 'clip_ratio/low_mean': 0.09098550211638212, 'clip_ratio/low_min': 0.09098550211638212, 'clip_ratio/high_mean': 0.11108248308300972, 'clip_ratio/high_max': 0.11108248308300972, 'clip_ratio/region_mean': 0.20206798519939184, 'reward_total_mean': 0.20144273340702057, 'reward_meter_mean': 0.48397544026374817, 'reward_meter_std': 0.2964073717594147, '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_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.9771234393119812, 'reward_repeat_penalty_std': 0.026555240154266357, 'reward_near_duplicate_penalty_mean': 0.9825642704963684, 'reward_near_duplicate_penalty_std': 0.014555198140442371, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9943438768386841, 'reward_distinct_2_std': 0.015997890383005142, 'reward_judge_quality_mean': 0.48374998569488525, 'reward_judge_quality_std': 0.18196842074394226, 'reward_total_composite_mean': 0.20144273340702057, 'reward_total_composite_std': 0.13822034001350403, 'epoch': 0.01}

  6%|▌         | 202/3300 [44:40<34:03:20, 39.57s/it]INFO 04-12 11:30:53 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 203/3300 [44:51<26:39:25, 30.99s/it]2026-04-12 11:31:03,799 | INFO | train_grpo_train | metrics_logged mode=train step=203

                                                     
{'loss': 0.0833, 'grad_norm': 13.165058135986328, 'learning_rate': 9.387878787878789e-06, 'num_tokens': 425960.0, 'completions/mean_length': 63.375, 'completions/min_length': 47.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.375, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.6267312169075012, 'rewards/meter/std': 0.348740816116333, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4624999761581421, 'rewards/judge_quality/std': 0.08345229178667068, 'rewards/repeat_penalty/mean': 0.9877314567565918, 'rewards/repeat_penalty/std': 0.01641863025724888, 'rewards/near_duplicate_penalty/mean': 0.9877314567565918, 'rewards/near_duplicate_penalty/std': 0.01641863025724888, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2830621600151062, 'rewards/total_composite/std': 0.1662304848432541, 'reward': 0.2830621600151062, 'reward_std': 0.1662304699420929, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22839851677417755, 'sampling/sampling_logp_difference/max': 1.754427433013916, 'sampling/importance_sampling_ratio/min': 0.17300626635551453, 'sampling/importance_sampling_ratio/mean': 1.0571361780166626, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.7351140081882477, 'clip_ratio/low_mean': 0.07908606342971325, 'clip_ratio/low_min': 0.07908606342971325, 'clip_ratio/high_mean': 0.14272558316588402, 'clip_ratio/high_max': 0.14272558316588402, 'clip_ratio/region_mean': 0.22181164659559727, 'reward_total_mean': 0.2830621600151062, 'reward_meter_mean': 0.6267312169075012, 'reward_meter_std': 0.348740816116333, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9877314567565918, 'reward_repeat_penalty_std': 0.01641863025724888, 'reward_near_duplicate_penalty_mean': 0.9877314567565918, 'reward_near_duplicate_penalty_std': 0.01641863025724888, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4624999761581421, 'reward_judge_quality_std': 0.08345229178667068, 'reward_total_composite_mean': 0.2830621600151062, 'reward_total_composite_std': 0.1662304848432541, 'epoch': 0.01}

  6%|▌         | 203/3300 [44:51<26:39:25, 30.99s/it]INFO 04-12 11:31:04 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 204/3300 [45:03<21:41:06, 25.22s/it]2026-04-12 11:31:15,546 | INFO | train_grpo_train | metrics_logged mode=train step=204

                                                     
{'loss': 0.086, 'grad_norm': 9.738399505615234, 'learning_rate': 9.384848484848486e-06, 'num_tokens': 428504.0, 'completions/mean_length': 118.0, 'completions/min_length': 98.0, 'completions/max_length': 177.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 118.0, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 177.0, 'rewards/meter/mean': 0.42173734307289124, 'rewards/meter/std': 0.3321233093738556, 'rewards/count_adherence/mean': 0.9791666269302368, 'rewards/count_adherence/std': 0.0589255727827549, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9744639992713928, 'rewards/lexical_plausibility/std': 0.04792870581150055, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.13024701178073883, 'rewards/repeat_penalty/mean': 0.9906821846961975, 'rewards/repeat_penalty/std': 0.011881274171173573, 'rewards/near_duplicate_penalty/mean': 0.9938464760780334, 'rewards/near_duplicate_penalty/std': 0.003660287708044052, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9967948794364929, 'rewards/distinct_2/std': 0.009065471589565277, 'rewards/total_composite/mean': 0.13917693495750427, 'rewards/total_composite/std': 0.1455865055322647, 'reward': 0.13917693495750427, 'reward_std': 0.1455865055322647, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22692568600177765, 'sampling/sampling_logp_difference/max': 1.8165950775146484, 'sampling/importance_sampling_ratio/min': 0.16257837414741516, 'sampling/importance_sampling_ratio/mean': 1.043134093284607, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.9794090688228607, 'clip_ratio/low_mean': 0.11748682335019112, 'clip_ratio/low_min': 0.11748682335019112, 'clip_ratio/high_mean': 0.09248615801334381, 'clip_ratio/high_max': 0.09248615801334381, 'clip_ratio/region_mean': 0.20997298136353493, 'reward_total_mean': 0.13917693495750427, 'reward_meter_mean': 0.42173734307289124, 'reward_meter_std': 0.3321233093738556, 'reward_count_adherence_mean': 0.9791666269302368, 'reward_count_adherence_std': 0.0589255727827549, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9744639992713928, 'reward_lexical_plausibility_std': 0.04792870581150055, 'reward_repeat_penalty_mean': 0.9906821846961975, 'reward_repeat_penalty_std': 0.011881274171173573, 'reward_near_duplicate_penalty_mean': 0.9938464760780334, 'reward_near_duplicate_penalty_std': 0.003660287708044052, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9967948794364929, 'reward_distinct_2_std': 0.009065471589565277, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.13024701178073883, 'reward_total_composite_mean': 0.13917693495750427, 'reward_total_composite_std': 0.1455865055322647, 'epoch': 0.01}

  6%|▌         | 204/3300 [45:03<21:41:06, 25.22s/it]INFO 04-12 11:31:15 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 205/3300 [45:13<17:50:59, 20.76s/it]2026-04-12 11:31:25,918 | INFO | train_grpo_train | metrics_logged mode=train step=205

                                                     
{'loss': -0.0142, 'grad_norm': 20.657299041748047, 'learning_rate': 9.381818181818183e-06, 'num_tokens': 429834.0, 'completions/mean_length': 19.25, 'completions/min_length': 14.0, 'completions/max_length': 23.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.25, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.2705998122692108, 'rewards/meter/std': 0.4059388041496277, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6725000143051147, 'rewards/judge_quality/std': 0.2666056752204895, 'rewards/repeat_penalty/mean': 0.7417961359024048, 'rewards/repeat_penalty/std': 0.019032426178455353, 'rewards/near_duplicate_penalty/mean': 0.9890615344047546, 'rewards/near_duplicate_penalty/std': 0.025376567617058754, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.13873904943466187, 'rewards/total_composite/std': 0.1813613921403885, 'reward': 0.13873904943466187, 'reward_std': 0.1813613772392273, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21271474659442902, 'sampling/sampling_logp_difference/max': 2.7158546447753906, 'sampling/importance_sampling_ratio/min': 0.06614840030670166, 'sampling/importance_sampling_ratio/mean': 1.0115232467651367, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6228943169116974, 'clip_ratio/low_mean': 0.11069627152755857, 'clip_ratio/low_min': 0.11069627152755857, 'clip_ratio/high_mean': 0.10267857275903225, 'clip_ratio/high_max': 0.10267857275903225, 'clip_ratio/region_mean': 0.21337484428659081, 'reward_total_mean': 0.13873904943466187, 'reward_meter_mean': 0.2705998122692108, 'reward_meter_std': 0.4059388041496277, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7417961359024048, 'reward_repeat_penalty_std': 0.019032426178455353, 'reward_near_duplicate_penalty_mean': 0.9890615344047546, 'reward_near_duplicate_penalty_std': 0.025376567617058754, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6725000143051147, 'reward_judge_quality_std': 0.2666056752204895, 'reward_total_composite_mean': 0.13873904943466187, 'reward_total_composite_std': 0.1813613921403885, 'epoch': 0.01}

  6%|▌         | 205/3300 [45:13<17:50:59, 20.76s/it]INFO 04-12 11:31:26 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 206/3300 [45:25<15:26:02, 17.96s/it]2026-04-12 11:31:37,333 | INFO | train_grpo_train | metrics_logged mode=train step=206

                                                     
{'loss': 0.0229, 'grad_norm': 9.159485816955566, 'learning_rate': 9.378787878787879e-06, 'num_tokens': 432181.0, 'completions/mean_length': 113.375, 'completions/min_length': 105.0, 'completions/max_length': 118.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 113.375, 'completions/min_terminated_length': 105.0, 'completions/max_terminated_length': 118.0, 'rewards/meter/mean': 0.5422508716583252, 'rewards/meter/std': 0.2832096219062805, 'rewards/count_adherence/mean': 0.9791666269302368, 'rewards/count_adherence/std': 0.0589255727827549, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9863553047180176, 'rewards/lexical_plausibility/std': 0.030255554243922234, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.994197428226471, 'rewards/repeat_penalty/std': 0.0035559621173888445, 'rewards/near_duplicate_penalty/mean': 0.994197428226471, 'rewards/near_duplicate_penalty/std': 0.0035559621173888445, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.19081418216228485, 'rewards/total_composite/std': 0.1504496932029724, 'reward': 0.19081418216228485, 'reward_std': 0.1504496932029724, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24501466751098633, 'sampling/sampling_logp_difference/max': 1.6380634307861328, 'sampling/importance_sampling_ratio/min': 0.1943560689687729, 'sampling/importance_sampling_ratio/mean': 1.0631917715072632, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.452824503183365, 'clip_ratio/low_mean': 0.08842740207910538, 'clip_ratio/low_min': 0.08842740207910538, 'clip_ratio/high_mean': 0.10499803721904755, 'clip_ratio/high_max': 0.10499803721904755, 'clip_ratio/region_mean': 0.19342543929815292, 'reward_total_mean': 0.19081418216228485, 'reward_meter_mean': 0.5422508716583252, 'reward_meter_std': 0.2832096219062805, 'reward_count_adherence_mean': 0.9791666269302368, 'reward_count_adherence_std': 0.0589255727827549, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9863553047180176, 'reward_lexical_plausibility_std': 0.030255554243922234, 'reward_repeat_penalty_mean': 0.994197428226471, 'reward_repeat_penalty_std': 0.0035559621173888445, 'reward_near_duplicate_penalty_mean': 0.994197428226471, 'reward_near_duplicate_penalty_std': 0.0035559621173888445, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.19081418216228485, 'reward_total_composite_std': 0.1504496932029724, 'epoch': 0.01}

  6%|▌         | 206/3300 [45:25<15:26:02, 17.96s/it]INFO 04-12 11:31:37 [block_pool.py:378] Successfully reset prefix cache

  6%|▋         | 207/3300 [45:34<13:14:26, 15.41s/it]2026-04-12 11:31:46,800 | INFO | train_grpo_train | metrics_logged mode=train step=207

                                                     
{'loss': 0.1455, 'grad_norm': 12.16246509552002, 'learning_rate': 9.375757575757576e-06, 'num_tokens': 434091.0, 'completions/mean_length': 65.75, 'completions/min_length': 34.0, 'completions/max_length': 94.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.75, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.2150837779045105, 'rewards/meter/std': 0.20138275623321533, 'rewards/count_adherence/mean': 0.84375, 'rewards/count_adherence/std': 0.18600596487522125, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.8963750600814819, 'rewards/lexical_plausibility/std': 0.15007846057415009, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.11952286213636398, 'rewards/repeat_penalty/mean': 0.9773169755935669, 'rewards/repeat_penalty/std': 0.028806151822209358, 'rewards/near_duplicate_penalty/mean': 0.9923085570335388, 'rewards/near_duplicate_penalty/std': 0.012803086079657078, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9847860336303711, 'rewards/distinct_2/std': 0.021188830956816673, 'rewards/total_composite/mean': 0.027725128456950188, 'rewards/total_composite/std': 0.023525994271039963, 'reward': 0.027725128456950188, 'reward_std': 0.023525994271039963, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2611384391784668, 'sampling/sampling_logp_difference/max': 1.3508033752441406, 'sampling/importance_sampling_ratio/min': 0.25903207063674927, 'sampling/importance_sampling_ratio/mean': 1.0614619255065918, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.64569291472435, 'clip_ratio/low_mean': 0.1289540696889162, 'clip_ratio/low_min': 0.1289540696889162, 'clip_ratio/high_mean': 0.07995899952948093, 'clip_ratio/high_max': 0.07995899952948093, 'clip_ratio/region_mean': 0.20891306921839714, 'reward_total_mean': 0.027725128456950188, 'reward_meter_mean': 0.2150837779045105, 'reward_meter_std': 0.20138275623321533, 'reward_count_adherence_mean': 0.84375, 'reward_count_adherence_std': 0.18600596487522125, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.8963750600814819, 'reward_lexical_plausibility_std': 0.15007846057415009, 'reward_repeat_penalty_mean': 0.9773169755935669, 'reward_repeat_penalty_std': 0.028806151822209358, 'reward_near_duplicate_penalty_mean': 0.9923085570335388, 'reward_near_duplicate_penalty_std': 0.012803086079657078, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9847860336303711, 'reward_distinct_2_std': 0.021188830956816673, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.11952286213636398, 'reward_total_composite_mean': 0.027725128456950188, 'reward_total_composite_std': 0.023525994271039963, 'epoch': 0.01}

  6%|▋         | 207/3300 [45:34<13:14:26, 15.41s/it]INFO 04-12 11:31:47 [block_pool.py:378] Successfully reset prefix cache

  6%|▋         | 208/3300 [45:44<11:53:58, 13.85s/it]2026-04-12 11:31:57,022 | INFO | train_grpo_train | metrics_logged mode=train step=208

                                                     
{'loss': 0.1053, 'grad_norm': 16.431283950805664, 'learning_rate': 9.372727272727273e-06, 'num_tokens': 435753.0, 'completions/mean_length': 38.75, 'completions/min_length': 31.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.75, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.5914057493209839, 'rewards/meter/std': 0.413548082113266, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.5550000071525574, 'rewards/judge_quality/std': 0.22790977358818054, 'rewards/repeat_penalty/mean': 0.9763509035110474, 'rewards/repeat_penalty/std': 0.01747816801071167, 'rewards/near_duplicate_penalty/mean': 0.9763509035110474, 'rewards/near_duplicate_penalty/std': 0.01747816801071167, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.33764782547950745, 'rewards/total_composite/std': 0.263327956199646, 'reward': 0.33764782547950745, 'reward_std': 0.263327956199646, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1850842535495758, 'sampling/sampling_logp_difference/max': 1.4391326904296875, 'sampling/importance_sampling_ratio/min': 0.23713333904743195, 'sampling/importance_sampling_ratio/mean': 1.0372931957244873, 'sampling/importance_sampling_ratio/max': 1.936633825302124, 'entropy': 1.8904611021280289, 'clip_ratio/low_mean': 0.08714842796325684, 'clip_ratio/low_min': 0.08714842796325684, 'clip_ratio/high_mean': 0.07172770705074072, 'clip_ratio/high_max': 0.07172770705074072, 'clip_ratio/region_mean': 0.15887613501399755, 'reward_total_mean': 0.33764782547950745, 'reward_meter_mean': 0.5914057493209839, 'reward_meter_std': 0.413548082113266, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.9763509035110474, 'reward_repeat_penalty_std': 0.01747816801071167, 'reward_near_duplicate_penalty_mean': 0.9763509035110474, 'reward_near_duplicate_penalty_std': 0.01747816801071167, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5550000071525574, 'reward_judge_quality_std': 0.22790977358818054, 'reward_total_composite_mean': 0.33764782547950745, 'reward_total_composite_std': 0.263327956199646, 'epoch': 0.01}

  6%|▋         | 208/3300 [45:44<11:53:58, 13.85s/it]INFO 04-12 11:31:57 [block_pool.py:378] Successfully reset prefix cache

  6%|▋         | 209/3300 [45:55<10:54:51, 12.71s/it]2026-04-12 11:32:07,069 | INFO | train_grpo_train | metrics_logged mode=train step=209

                                                     
{'loss': 0.0735, 'grad_norm': 14.297948837280273, 'learning_rate': 9.36969696969697e-06, 'num_tokens': 437344.0, 'completions/mean_length': 36.875, 'completions/min_length': 30.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.875, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.17486268281936646, 'rewards/meter/std': 0.34384360909461975, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6012499928474426, 'rewards/judge_quality/std': 0.2671777307987213, 'rewards/repeat_penalty/mean': 0.9022783637046814, 'rewards/repeat_penalty/std': 0.12628450989723206, 'rewards/near_duplicate_penalty/mean': 0.9960283637046814, 'rewards/near_duplicate_penalty/std': 0.007600844372063875, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.06744881719350815, 'rewards/total_composite/std': 0.11959356814622879, 'reward': 0.06744881719350815, 'reward_std': 0.11959356814622879, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21043027937412262, 'sampling/sampling_logp_difference/max': 1.7506370544433594, 'sampling/importance_sampling_ratio/min': 0.17366325855255127, 'sampling/importance_sampling_ratio/mean': 1.046602725982666, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.123324006795883, 'clip_ratio/low_mean': 0.08261993806809187, 'clip_ratio/low_min': 0.08261993806809187, 'clip_ratio/high_mean': 0.0787318553775549, 'clip_ratio/high_max': 0.0787318553775549, 'clip_ratio/region_mean': 0.16135179344564676, 'reward_total_mean': 0.06744881719350815, 'reward_meter_mean': 0.17486268281936646, 'reward_meter_std': 0.34384360909461975, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9022783637046814, 'reward_repeat_penalty_std': 0.12628450989723206, 'reward_near_duplicate_penalty_mean': 0.9960283637046814, 'reward_near_duplicate_penalty_std': 0.007600844372063875, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6012499928474426, 'reward_judge_quality_std': 0.2671777307987213, 'reward_total_composite_mean': 0.06744881719350815, 'reward_total_composite_std': 0.11959356814622879, 'epoch': 0.01}

  6%|▋         | 209/3300 [45:55<10:54:51, 12.71s/it]INFO 04-12 11:32:07 [block_pool.py:378] Successfully reset prefix cache

  6%|▋         | 210/3300 [46:06<10:29:36, 12.23s/it]2026-04-12 11:32:18,160 | INFO | train_grpo_train | metrics_logged mode=train step=210

                                                     
{'loss': 0.0118, 'grad_norm': 11.766057014465332, 'learning_rate': 9.366666666666668e-06, 'num_tokens': 439282.0, 'completions/mean_length': 70.25, 'completions/min_length': 61.0, 'completions/max_length': 89.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.25, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 89.0, 'rewards/meter/mean': 0.28878700733184814, 'rewards/meter/std': 0.4075119197368622, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.13093073666095734, 'rewards/repeat_penalty/mean': 0.9284815192222595, 'rewards/repeat_penalty/std': 0.14472460746765137, 'rewards/near_duplicate_penalty/mean': 0.9761557579040527, 'rewards/near_duplicate_penalty/std': 0.02489008754491806, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9613729119300842, 'rewards/distinct_2/std': 0.0950038805603981, 'rewards/total_composite/mean': 0.06665698438882828, 'rewards/total_composite/std': 0.13737748563289642, 'reward': 0.06665698438882828, 'reward_std': 0.13737748563289642, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23840616643428802, 'sampling/sampling_logp_difference/max': 1.4764100313186646, 'sampling/importance_sampling_ratio/min': 0.22845637798309326, 'sampling/importance_sampling_ratio/mean': 1.0464091300964355, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.9943851232528687, 'clip_ratio/low_mean': 0.1300221085548401, 'clip_ratio/low_min': 0.1300221085548401, 'clip_ratio/high_mean': 0.0685396883636713, 'clip_ratio/high_max': 0.0685396883636713, 'clip_ratio/region_mean': 0.1985617969185114, 'reward_total_mean': 0.06665698438882828, 'reward_meter_mean': 0.28878700733184814, 'reward_meter_std': 0.4075119197368622, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9284815192222595, 'reward_repeat_penalty_std': 0.14472460746765137, 'reward_near_duplicate_penalty_mean': 0.9761557579040527, 'reward_near_duplicate_penalty_std': 0.02489008754491806, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9613729119300842, 'reward_distinct_2_std': 0.0950038805603981, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.13093073666095734, 'reward_total_composite_mean': 0.06665698438882828, 'reward_total_composite_std': 0.13737748563289642, 'epoch': 0.01}

  6%|▋         | 210/3300 [46:06<10:29:36, 12.23s/it]INFO 04-12 11:32:18 [block_pool.py:378] Successfully reset prefix cache

  6%|▋         | 211/3300 [46:16<10:03:19, 11.72s/it]2026-04-12 11:32:28,695 | INFO | train_grpo_train | metrics_logged mode=train step=211

                                                     
{'loss': -0.0371, 'grad_norm': 10.741786003112793, 'learning_rate': 9.363636363636365e-06, 'num_tokens': 441558.0, 'completions/mean_length': 87.5, 'completions/min_length': 77.0, 'completions/max_length': 100.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 87.5, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 100.0, 'rewards/meter/mean': 0.5041914582252502, 'rewards/meter/std': 0.3094952702522278, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.18600596487522125, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9887020587921143, 'rewards/lexical_plausibility/std': 0.025517085567116737, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.14078859984874725, 'rewards/repeat_penalty/mean': 0.9968490600585938, 'rewards/repeat_penalty/std': 0.0037891643587499857, 'rewards/near_duplicate_penalty/mean': 0.9968490600585938, 'rewards/near_duplicate_penalty/std': 0.0037891643587499857, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.13640481233596802, 'rewards/total_composite/std': 0.1014382466673851, 'reward': 0.13640481233596802, 'reward_std': 0.1014382466673851, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2291981279850006, 'sampling/sampling_logp_difference/max': 1.4337005615234375, 'sampling/importance_sampling_ratio/min': 0.23842500150203705, 'sampling/importance_sampling_ratio/mean': 1.0694539546966553, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.521673172712326, 'clip_ratio/low_mean': 0.1282580867409706, 'clip_ratio/low_min': 0.1282580867409706, 'clip_ratio/high_mean': 0.080580810084939, 'clip_ratio/high_max': 0.080580810084939, 'clip_ratio/region_mean': 0.20883889682590961, 'reward_total_mean': 0.13640481233596802, 'reward_meter_mean': 0.5041914582252502, 'reward_meter_std': 0.3094952702522278, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.18600596487522125, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9887020587921143, 'reward_lexical_plausibility_std': 0.025517085567116737, 'reward_repeat_penalty_mean': 0.9968490600585938, 'reward_repeat_penalty_std': 0.0037891643587499857, 'reward_near_duplicate_penalty_mean': 0.9968490600585938, 'reward_near_duplicate_penalty_std': 0.0037891643587499857, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.14078859984874725, 'reward_total_composite_mean': 0.13640481233596802, 'reward_total_composite_std': 0.1014382466673851, 'epoch': 0.01}

  6%|▋         | 211/3300 [46:16<10:03:19, 11.72s/it]INFO 04-12 11:32:28 [block_pool.py:378] Successfully reset prefix cache

  6%|▋         | 212/3300 [46:25<9:21:17, 10.91s/it] 2026-04-12 11:32:37,708 | INFO | train_grpo_train | metrics_logged mode=train step=212

                                                    
{'loss': 0.0301, 'grad_norm': 18.575624465942383, 'learning_rate': 9.36060606060606e-06, 'num_tokens': 443146.0, 'completions/mean_length': 38.5, 'completions/min_length': 33.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.5, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.5946434736251831, 'rewards/meter/std': 0.4882367253303528, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6850000023841858, 'rewards/judge_quality/std': 0.2512255907058716, 'rewards/repeat_penalty/mean': 0.9368149638175964, 'rewards/repeat_penalty/std': 0.11532019823789597, 'rewards/near_duplicate_penalty/mean': 0.9993149638175964, 'rewards/near_duplicate_penalty/std': 0.0019375960109755397, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.49817895889282227, 'rewards/total_composite/std': 0.448115736246109, 'reward': 0.49817895889282227, 'reward_std': 0.44811567664146423, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19288630783557892, 'sampling/sampling_logp_difference/max': 1.7310380935668945, 'sampling/importance_sampling_ratio/min': 0.17710046470165253, 'sampling/importance_sampling_ratio/mean': 1.0244054794311523, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.718473881483078, 'clip_ratio/low_mean': 0.10665038600564003, 'clip_ratio/low_min': 0.10665038600564003, 'clip_ratio/high_mean': 0.07521865144371986, 'clip_ratio/high_max': 0.07521865144371986, 'clip_ratio/region_mean': 0.1818690374493599, 'reward_total_mean': 0.49817895889282227, 'reward_meter_mean': 0.5946434736251831, 'reward_meter_std': 0.4882367253303528, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9368149638175964, 'reward_repeat_penalty_std': 0.11532019823789597, 'reward_near_duplicate_penalty_mean': 0.9993149638175964, 'reward_near_duplicate_penalty_std': 0.0019375960109755397, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6850000023841858, 'reward_judge_quality_std': 0.2512255907058716, 'reward_total_composite_mean': 0.49817895889282227, 'reward_total_composite_std': 0.448115736246109, 'epoch': 0.01}

  6%|▋         | 212/3300 [46:25<9:21:17, 10.91s/it]INFO 04-12 11:32:37 [block_pool.py:378] Successfully reset prefix cache

  6%|▋         | 213/3300 [46:36<9:20:52, 10.90s/it]2026-04-12 11:32:48,595 | INFO | train_grpo_train | metrics_logged mode=train step=213

                                                    
{'loss': -0.0341, 'grad_norm': 13.296220779418945, 'learning_rate': 9.357575757575758e-06, 'num_tokens': 445305.0, 'completions/mean_length': 76.875, 'completions/min_length': 66.0, 'completions/max_length': 94.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.875, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.7125569581985474, 'rewards/meter/std': 0.31757718324661255, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4375, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.9858448505401611, 'rewards/repeat_penalty/std': 0.01981234923005104, 'rewards/near_duplicate_penalty/mean': 0.9906812906265259, 'rewards/near_duplicate_penalty/std': 0.009823467582464218, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9950690269470215, 'rewards/distinct_2/std': 0.013946877792477608, 'rewards/total_composite/mean': 0.3205989599227905, 'rewards/total_composite/std': 0.16013184189796448, 'reward': 0.3205989599227905, 'reward_std': 0.16013184189796448, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23128631711006165, 'sampling/sampling_logp_difference/max': 1.755228042602539, 'sampling/importance_sampling_ratio/min': 0.1728678196668625, 'sampling/importance_sampling_ratio/mean': 1.0455572605133057, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6935992538928986, 'clip_ratio/low_mean': 0.0612958949059248, 'clip_ratio/low_min': 0.0612958949059248, 'clip_ratio/high_mean': 0.13209493644535542, 'clip_ratio/high_max': 0.13209493644535542, 'clip_ratio/region_mean': 0.1933908313512802, 'reward_total_mean': 0.3205989599227905, 'reward_meter_mean': 0.7125569581985474, 'reward_meter_std': 0.31757718324661255, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9858448505401611, 'reward_repeat_penalty_std': 0.01981234923005104, 'reward_near_duplicate_penalty_mean': 0.9906812906265259, 'reward_near_duplicate_penalty_std': 0.009823467582464218, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9950690269470215, 'reward_distinct_2_std': 0.013946877792477608, 'reward_judge_quality_mean': 0.4375, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.3205989599227905, 'reward_total_composite_std': 0.16013184189796448, 'epoch': 0.01}

  6%|▋         | 213/3300 [46:36<9:20:52, 10.90s/it]INFO 04-12 11:32:48 [block_pool.py:378] Successfully reset prefix cache

  6%|▋         | 214/3300 [46:47<9:20:06, 10.89s/it]2026-04-12 11:32:59,459 | INFO | train_grpo_train | metrics_logged mode=train step=214

                                                    
{'loss': -0.1236, 'grad_norm': 17.280521392822266, 'learning_rate': 9.354545454545455e-06, 'num_tokens': 446842.0, 'completions/mean_length': 28.125, 'completions/min_length': 19.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.125, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.4867746829986572, 'rewards/meter/std': 0.34936103224754333, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.821280837059021, 'rewards/lexical_plausibility/std': 0.22418732941150665, 'rewards/judge_quality/mean': 0.4462500214576721, 'rewards/judge_quality/std': 0.40896865725517273, 'rewards/repeat_penalty/mean': 0.7469173073768616, 'rewards/repeat_penalty/std': 0.00871915090829134, 'rewards/near_duplicate_penalty/mean': 0.9958897829055786, 'rewards/near_duplicate_penalty/std': 0.011625555343925953, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.21777765452861786, 'rewards/total_composite/std': 0.30124431848526, 'reward': 0.21777765452861786, 'reward_std': 0.30124431848526, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19614949822425842, 'sampling/sampling_logp_difference/max': 1.2957849502563477, 'sampling/importance_sampling_ratio/min': 0.27368295192718506, 'sampling/importance_sampling_ratio/mean': 1.036525011062622, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5247266441583633, 'clip_ratio/low_mean': 0.15370856132358313, 'clip_ratio/low_min': 0.15370856132358313, 'clip_ratio/high_mean': 0.053542179986834526, 'clip_ratio/high_max': 0.053542179986834526, 'clip_ratio/region_mean': 0.20725074131041765, 'reward_total_mean': 0.21777765452861786, 'reward_meter_mean': 0.4867746829986572, 'reward_meter_std': 0.34936103224754333, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.821280837059021, 'reward_lexical_plausibility_std': 0.22418732941150665, 'reward_repeat_penalty_mean': 0.7469173073768616, 'reward_repeat_penalty_std': 0.00871915090829134, 'reward_near_duplicate_penalty_mean': 0.9958897829055786, 'reward_near_duplicate_penalty_std': 0.011625555343925953, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4462500214576721, 'reward_judge_quality_std': 0.40896865725517273, 'reward_total_composite_mean': 0.21777765452861786, 'reward_total_composite_std': 0.30124431848526, 'epoch': 0.01}

  6%|▋         | 214/3300 [46:47<9:20:06, 10.89s/it]INFO 04-12 11:32:59 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 215/3300 [46:57<9:00:34, 10.51s/it]2026-04-12 11:33:09,095 | INFO | train_grpo_train | metrics_logged mode=train step=215

                                                    
{'loss': -0.0037, 'grad_norm': 18.812461853027344, 'learning_rate': 9.351515151515152e-06, 'num_tokens': 448296.0, 'completions/mean_length': 32.75, 'completions/min_length': 27.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.75, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.7190053462982178, 'rewards/meter/std': 0.38506612181663513, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6012499928474426, 'rewards/judge_quality/std': 0.2671777307987213, 'rewards/repeat_penalty/mean': 0.9549033641815186, 'rewards/repeat_penalty/std': 0.06095125898718834, 'rewards/near_duplicate_penalty/mean': 0.9763610363006592, 'rewards/near_duplicate_penalty/std': 0.02465181238949299, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9772970080375671, 'rewards/distinct_2/std': 0.04213463515043259, 'rewards/total_composite/mean': 0.4101389944553375, 'rewards/total_composite/std': 0.2677665948867798, 'reward': 0.4101389944553375, 'reward_std': 0.2677665948867798, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18426816165447235, 'sampling/sampling_logp_difference/max': 1.6734638214111328, 'sampling/importance_sampling_ratio/min': 0.1875961273908615, 'sampling/importance_sampling_ratio/mean': 1.0101020336151123, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3771517127752304, 'clip_ratio/low_mean': 0.06315710954368114, 'clip_ratio/low_min': 0.06315710954368114, 'clip_ratio/high_mean': 0.07150485645979643, 'clip_ratio/high_max': 0.07150485645979643, 'clip_ratio/region_mean': 0.13466196600347757, 'reward_total_mean': 0.4101389944553375, 'reward_meter_mean': 0.7190053462982178, 'reward_meter_std': 0.38506612181663513, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9549033641815186, 'reward_repeat_penalty_std': 0.06095125898718834, 'reward_near_duplicate_penalty_mean': 0.9763610363006592, 'reward_near_duplicate_penalty_std': 0.02465181238949299, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9772970080375671, 'reward_distinct_2_std': 0.04213463515043259, 'reward_judge_quality_mean': 0.6012499928474426, 'reward_judge_quality_std': 0.2671777307987213, 'reward_total_composite_mean': 0.4101389944553375, 'reward_total_composite_std': 0.2677665948867798, 'epoch': 0.01}

  7%|▋         | 215/3300 [46:57<9:00:34, 10.51s/it]INFO 04-12 11:33:09 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 216/3300 [47:05<8:23:10,  9.79s/it]2026-04-12 11:33:17,194 | INFO | train_grpo_train | metrics_logged mode=train step=216

                                                    
{'loss': -0.0106, 'grad_norm': 22.502124786376953, 'learning_rate': 9.34848484848485e-06, 'num_tokens': 449620.0, 'completions/mean_length': 17.5, 'completions/min_length': 14.0, 'completions/max_length': 25.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 17.5, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.9235381484031677, 'rewards/meter/std': 0.10606560856103897, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7762500047683716, 'rewards/judge_quality/std': 0.30500292778015137, 'rewards/repeat_penalty/mean': 0.7335008978843689, 'rewards/repeat_penalty/std': 0.04226246103644371, 'rewards/near_duplicate_penalty/mean': 0.9780011773109436, 'rewards/near_duplicate_penalty/std': 0.05634994059801102, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.7379944920539856, 'rewards/total_composite/std': 0.302329957485199, 'reward': 0.7379944920539856, 'reward_std': 0.302329957485199, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17037364840507507, 'sampling/sampling_logp_difference/max': 1.2017364501953125, 'sampling/importance_sampling_ratio/min': 0.30067166686058044, 'sampling/importance_sampling_ratio/mean': 1.0505744218826294, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4617980793118477, 'clip_ratio/low_mean': 0.0546875, 'clip_ratio/low_min': 0.0546875, 'clip_ratio/high_mean': 0.07439980236813426, 'clip_ratio/high_max': 0.07439980236813426, 'clip_ratio/region_mean': 0.12908730236813426, 'reward_total_mean': 0.7379944920539856, 'reward_meter_mean': 0.9235381484031677, 'reward_meter_std': 0.10606560856103897, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7335008978843689, 'reward_repeat_penalty_std': 0.04226246103644371, 'reward_near_duplicate_penalty_mean': 0.9780011773109436, 'reward_near_duplicate_penalty_std': 0.05634994059801102, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7762500047683716, 'reward_judge_quality_std': 0.30500292778015137, 'reward_total_composite_mean': 0.7379944920539856, 'reward_total_composite_std': 0.302329957485199, 'epoch': 0.01}

  7%|▋         | 216/3300 [47:05<8:23:10,  9.79s/it]INFO 04-12 11:33:17 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 217/3300 [47:14<8:15:27,  9.64s/it]2026-04-12 11:33:26,494 | INFO | train_grpo_train | metrics_logged mode=train step=217

                                                    
{'loss': 0.1211, 'grad_norm': 23.88068199157715, 'learning_rate': 9.345454545454547e-06, 'num_tokens': 451255.0, 'completions/mean_length': 40.375, 'completions/min_length': 24.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.375, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.3083277940750122, 'rewards/meter/std': 0.33922672271728516, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.936546802520752, 'rewards/lexical_plausibility/std': 0.1613805890083313, 'rewards/judge_quality/mean': 0.39625000953674316, 'rewards/judge_quality/std': 0.24136146903038025, 'rewards/repeat_penalty/mean': 0.9800436496734619, 'rewards/repeat_penalty/std': 0.03134879842400551, 'rewards/near_duplicate_penalty/mean': 0.9800436496734619, 'rewards/near_duplicate_penalty/std': 0.03134879842400551, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1443057507276535, 'rewards/total_composite/std': 0.19835418462753296, 'reward': 0.1443057507276535, 'reward_std': 0.19835416972637177, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23465116322040558, 'sampling/sampling_logp_difference/max': 1.6458568572998047, 'sampling/importance_sampling_ratio/min': 0.19284723699092865, 'sampling/importance_sampling_ratio/mean': 1.0250104665756226, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7267180606722832, 'clip_ratio/low_mean': 0.17233465798199177, 'clip_ratio/low_min': 0.17233465798199177, 'clip_ratio/high_mean': 0.04647871479392052, 'clip_ratio/high_max': 0.04647871479392052, 'clip_ratio/region_mean': 0.21881337277591228, 'reward_total_mean': 0.1443057507276535, 'reward_meter_mean': 0.3083277940750122, 'reward_meter_std': 0.33922672271728516, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.936546802520752, 'reward_lexical_plausibility_std': 0.1613805890083313, 'reward_repeat_penalty_mean': 0.9800436496734619, 'reward_repeat_penalty_std': 0.03134879842400551, 'reward_near_duplicate_penalty_mean': 0.9800436496734619, 'reward_near_duplicate_penalty_std': 0.03134879842400551, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.39625000953674316, 'reward_judge_quality_std': 0.24136146903038025, 'reward_total_composite_mean': 0.1443057507276535, 'reward_total_composite_std': 0.19835418462753296, 'epoch': 0.01}

  7%|▋         | 217/3300 [47:14<8:15:27,  9.64s/it]INFO 04-12 11:33:26 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 218/3300 [47:28<9:28:54, 11.08s/it]2026-04-12 11:33:40,912 | INFO | train_grpo_train | metrics_logged mode=train step=218

                                                    
{'loss': -0.03, 'grad_norm': 5.2542500495910645, 'learning_rate': 9.342424242424243e-06, 'num_tokens': 452699.0, 'completions/mean_length': 88.5, 'completions/min_length': 25.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 28.000001907348633, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.39811187982559204, 'rewards/meter/std': 0.47108978033065796, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.7219630479812622, 'rewards/lexical_plausibility/std': 0.3385985791683197, 'rewards/judge_quality/mean': 0.5637500286102295, 'rewards/judge_quality/std': 0.4051785171031952, 'rewards/repeat_penalty/mean': 0.7355536818504333, 'rewards/repeat_penalty/std': 0.14165730774402618, 'rewards/near_duplicate_penalty/mean': 0.939071536064148, 'rewards/near_duplicate_penalty/std': 0.1264227032661438, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2630048096179962, 'rewards/total_composite/std': 0.38137587904930115, 'reward': 0.2630048096179962, 'reward_std': 0.38137587904930115, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19810828566551208, 'sampling/sampling_logp_difference/max': 1.0608623027801514, 'sampling/importance_sampling_ratio/min': 0.3461571931838989, 'sampling/importance_sampling_ratio/mean': 1.0527364015579224, 'sampling/importance_sampling_ratio/max': 1.7826459407806396, 'entropy': 1.8758389204740524, 'clip_ratio/low_mean': 0.12835511937737465, 'clip_ratio/low_min': 0.12835511937737465, 'clip_ratio/high_mean': 0.047413794323801994, 'clip_ratio/high_max': 0.047413794323801994, 'clip_ratio/region_mean': 0.17576891370117664, 'reward_total_mean': 0.2630048096179962, 'reward_meter_mean': 0.39811187982559204, 'reward_meter_std': 0.47108978033065796, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.7219630479812622, 'reward_lexical_plausibility_std': 0.3385985791683197, 'reward_repeat_penalty_mean': 0.7355536818504333, 'reward_repeat_penalty_std': 0.14165730774402618, 'reward_near_duplicate_penalty_mean': 0.939071536064148, 'reward_near_duplicate_penalty_std': 0.1264227032661438, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5637500286102295, 'reward_judge_quality_std': 0.4051785171031952, 'reward_total_composite_mean': 0.2630048096179962, 'reward_total_composite_std': 0.38137587904930115, 'epoch': 0.01}

  7%|▋         | 218/3300 [47:28<9:28:54, 11.08s/it]INFO 04-12 11:33:41 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 219/3300 [47:38<9:13:38, 10.78s/it]2026-04-12 11:33:51,010 | INFO | train_grpo_train | metrics_logged mode=train step=219

                                                    
{'loss': 0.0847, 'grad_norm': 14.841551780700684, 'learning_rate': 9.33939393939394e-06, 'num_tokens': 454143.0, 'completions/mean_length': 38.5, 'completions/min_length': 32.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.5, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.7012842297554016, 'rewards/meter/std': 0.4026435315608978, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.8919566869735718, 'rewards/lexical_plausibility/std': 0.11086653172969818, 'rewards/judge_quality/mean': 0.23750001192092896, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.9982781410217285, 'rewards/repeat_penalty/std': 0.004219908732920885, 'rewards/near_duplicate_penalty/mean': 0.9982781410217285, 'rewards/near_duplicate_penalty/std': 0.004219908732920885, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.16786515712738037, 'rewards/total_composite/std': 0.1504647135734558, 'reward': 0.16786515712738037, 'reward_std': 0.1504647135734558, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2631381154060364, 'sampling/sampling_logp_difference/max': 1.5028114318847656, 'sampling/importance_sampling_ratio/min': 0.22250373661518097, 'sampling/importance_sampling_ratio/mean': 1.0572354793548584, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.5383233726024628, 'clip_ratio/low_mean': 0.15453205537050962, 'clip_ratio/low_min': 0.15453205537050962, 'clip_ratio/high_mean': 0.0546875, 'clip_ratio/high_max': 0.0546875, 'clip_ratio/region_mean': 0.20921955537050962, 'reward_total_mean': 0.16786515712738037, 'reward_meter_mean': 0.7012842297554016, 'reward_meter_std': 0.4026435315608978, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.8919566869735718, 'reward_lexical_plausibility_std': 0.11086653172969818, 'reward_repeat_penalty_mean': 0.9982781410217285, 'reward_repeat_penalty_std': 0.004219908732920885, 'reward_near_duplicate_penalty_mean': 0.9982781410217285, 'reward_near_duplicate_penalty_std': 0.004219908732920885, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.23750001192092896, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.16786515712738037, 'reward_total_composite_std': 0.1504647135734558, 'epoch': 0.01}

  7%|▋         | 219/3300 [47:38<9:13:38, 10.78s/it]INFO 04-12 11:33:51 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 220/3300 [47:49<9:06:37, 10.65s/it]2026-04-12 11:34:01,355 | INFO | train_grpo_train | metrics_logged mode=train step=220

                                                    
{'loss': 0.0005, 'grad_norm': 18.605060577392578, 'learning_rate': 9.336363636363637e-06, 'num_tokens': 455853.0, 'completions/mean_length': 33.75, 'completions/min_length': 29.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.75, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.6623882055282593, 'rewards/meter/std': 0.25034192204475403, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6475000381469727, 'rewards/judge_quality/std': 0.30742016434669495, 'rewards/repeat_penalty/mean': 0.9004184007644653, 'rewards/repeat_penalty/std': 0.11043259501457214, 'rewards/near_duplicate_penalty/mean': 0.9357359409332275, 'rewards/near_duplicate_penalty/std': 0.0423009991645813, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9595141410827637, 'rewards/distinct_2/std': 0.08656234294176102, 'rewards/total_composite/mean': 0.4125654101371765, 'rewards/total_composite/std': 0.2729375958442688, 'reward': 0.4125654101371765, 'reward_std': 0.2729375958442688, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18542027473449707, 'sampling/sampling_logp_difference/max': 1.2532129287719727, 'sampling/importance_sampling_ratio/min': 0.28558576107025146, 'sampling/importance_sampling_ratio/mean': 1.0390307903289795, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.536943256855011, 'clip_ratio/low_mean': 0.11440100893378258, 'clip_ratio/low_min': 0.11440100893378258, 'clip_ratio/high_mean': 0.05537925707176328, 'clip_ratio/high_max': 0.05537925707176328, 'clip_ratio/region_mean': 0.16978026600554585, 'reward_total_mean': 0.4125654101371765, 'reward_meter_mean': 0.6623882055282593, 'reward_meter_std': 0.25034192204475403, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9004184007644653, 'reward_repeat_penalty_std': 0.11043259501457214, 'reward_near_duplicate_penalty_mean': 0.9357359409332275, 'reward_near_duplicate_penalty_std': 0.0423009991645813, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9595141410827637, 'reward_distinct_2_std': 0.08656234294176102, 'reward_judge_quality_mean': 0.6475000381469727, 'reward_judge_quality_std': 0.30742016434669495, 'reward_total_composite_mean': 0.4125654101371765, 'reward_total_composite_std': 0.2729375958442688, 'epoch': 0.01}

  7%|▋         | 220/3300 [47:49<9:06:37, 10.65s/it]INFO 04-12 11:34:01 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 221/3300 [48:00<9:16:35, 10.85s/it]2026-04-12 11:34:12,666 | INFO | train_grpo_train | metrics_logged mode=train step=221

                                                    
{'loss': 0.0028, 'grad_norm': 15.065362930297852, 'learning_rate': 9.333333333333334e-06, 'num_tokens': 457426.0, 'completions/mean_length': 39.625, 'completions/min_length': 33.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.625, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.7895299196243286, 'rewards/meter/std': 0.30216193199157715, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9877232313156128, 'rewards/lexical_plausibility/std': 0.0218378733843565, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.9831031560897827, 'rewards/repeat_penalty/std': 0.0362558476626873, 'rewards/near_duplicate_penalty/mean': 0.9915370941162109, 'rewards/near_duplicate_penalty/std': 0.015189937315881252, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9912587404251099, 'rewards/distinct_2/std': 0.024724015966057777, 'rewards/total_composite/mean': 0.2931957542896271, 'rewards/total_composite/std': 0.13758660852909088, 'reward': 0.2931957542896271, 'reward_std': 0.13758662343025208, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21733666956424713, 'sampling/sampling_logp_difference/max': 1.1337356567382812, 'sampling/importance_sampling_ratio/min': 0.3218287527561188, 'sampling/importance_sampling_ratio/mean': 1.0576497316360474, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.408338636159897, 'clip_ratio/low_mean': 0.07987845689058304, 'clip_ratio/low_min': 0.07987845689058304, 'clip_ratio/high_mean': 0.12823371216654778, 'clip_ratio/high_max': 0.12823371216654778, 'clip_ratio/region_mean': 0.2081121690571308, 'reward_total_mean': 0.2931957542896271, 'reward_meter_mean': 0.7895299196243286, 'reward_meter_std': 0.30216193199157715, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9877232313156128, 'reward_lexical_plausibility_std': 0.0218378733843565, 'reward_repeat_penalty_mean': 0.9831031560897827, 'reward_repeat_penalty_std': 0.0362558476626873, 'reward_near_duplicate_penalty_mean': 0.9915370941162109, 'reward_near_duplicate_penalty_std': 0.015189937315881252, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9912587404251099, 'reward_distinct_2_std': 0.024724015966057777, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.2931957542896271, 'reward_total_composite_std': 0.13758660852909088, 'epoch': 0.01}

  7%|▋         | 221/3300 [48:00<9:16:35, 10.85s/it]INFO 04-12 11:34:13 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 222/3300 [48:10<9:04:18, 10.61s/it]2026-04-12 11:34:22,718 | INFO | train_grpo_train | metrics_logged mode=train step=222

                                                    
{'loss': 0.1503, 'grad_norm': 15.068031311035156, 'learning_rate': 9.33030303030303e-06, 'num_tokens': 459061.0, 'completions/mean_length': 43.375, 'completions/min_length': 31.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.375, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.45734867453575134, 'rewards/meter/std': 0.46583592891693115, '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/lexical_plausibility/mean': 0.9592617154121399, 'rewards/lexical_plausibility/std': 0.10445455461740494, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.9037517309188843, 'rewards/repeat_penalty/std': 0.17451131343841553, 'rewards/near_duplicate_penalty/mean': 0.9594180583953857, 'rewards/near_duplicate_penalty/std': 0.06871283054351807, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9622253179550171, 'rewards/distinct_2/std': 0.07333158701658249, 'rewards/total_composite/mean': 0.1565871685743332, 'rewards/total_composite/std': 0.20494528114795685, 'reward': 0.1565871685743332, 'reward_std': 0.20494528114795685, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23400035500526428, 'sampling/sampling_logp_difference/max': 2.14304780960083, 'sampling/importance_sampling_ratio/min': 0.11729680001735687, 'sampling/importance_sampling_ratio/mean': 1.0372645854949951, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6688247323036194, 'clip_ratio/low_mean': 0.0942892930470407, 'clip_ratio/low_min': 0.0942892930470407, 'clip_ratio/high_mean': 0.08344548381865025, 'clip_ratio/high_max': 0.08344548381865025, 'clip_ratio/region_mean': 0.17773477686569095, 'reward_total_mean': 0.1565871685743332, 'reward_meter_mean': 0.45734867453575134, 'reward_meter_std': 0.46583592891693115, '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_lexical_plausibility_mean': 0.9592617154121399, 'reward_lexical_plausibility_std': 0.10445455461740494, 'reward_repeat_penalty_mean': 0.9037517309188843, 'reward_repeat_penalty_std': 0.17451131343841553, 'reward_near_duplicate_penalty_mean': 0.9594180583953857, 'reward_near_duplicate_penalty_std': 0.06871283054351807, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9622253179550171, 'reward_distinct_2_std': 0.07333158701658249, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.1565871685743332, 'reward_total_composite_std': 0.20494528114795685, 'epoch': 0.01}

  7%|▋         | 222/3300 [48:10<9:04:18, 10.61s/it]INFO 04-12 11:34:22 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 223/3300 [48:20<8:59:22, 10.52s/it]2026-04-12 11:34:33,015 | INFO | train_grpo_train | metrics_logged mode=train step=223

                                                    
{'loss': -0.0877, 'grad_norm': 13.039738655090332, 'learning_rate': 9.327272727272729e-06, 'num_tokens': 460900.0, 'completions/mean_length': 55.875, 'completions/min_length': 47.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.875, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.1726960837841034, 'rewards/meter/std': 0.2386988401412964, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9042977094650269, 'rewards/lexical_plausibility/std': 0.2067929059267044, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.13562026619911194, 'rewards/repeat_penalty/mean': 0.9787956476211548, 'rewards/repeat_penalty/std': 0.03732248768210411, 'rewards/near_duplicate_penalty/mean': 0.9852607250213623, 'rewards/near_duplicate_penalty/std': 0.020230863243341446, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9931318759918213, 'rewards/distinct_2/std': 0.01942601054906845, 'rewards/total_composite/mean': 0.05157750844955444, 'rewards/total_composite/std': 0.0837271586060524, 'reward': 0.05157750844955444, 'reward_std': 0.0837271437048912, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23287220299243927, 'sampling/sampling_logp_difference/max': 1.4589438438415527, 'sampling/importance_sampling_ratio/min': 0.24223922193050385, 'sampling/importance_sampling_ratio/mean': 1.0731515884399414, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.2244983911514282, 'clip_ratio/low_mean': 0.15650768764317036, 'clip_ratio/low_min': 0.15650768764317036, 'clip_ratio/high_mean': 0.05032467655837536, 'clip_ratio/high_max': 0.05032467655837536, 'clip_ratio/region_mean': 0.20683236420154572, 'reward_total_mean': 0.05157750844955444, 'reward_meter_mean': 0.1726960837841034, 'reward_meter_std': 0.2386988401412964, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9042977094650269, 'reward_lexical_plausibility_std': 0.2067929059267044, 'reward_repeat_penalty_mean': 0.9787956476211548, 'reward_repeat_penalty_std': 0.03732248768210411, 'reward_near_duplicate_penalty_mean': 0.9852607250213623, 'reward_near_duplicate_penalty_std': 0.020230863243341446, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9931318759918213, 'reward_distinct_2_std': 0.01942601054906845, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.13562026619911194, 'reward_total_composite_mean': 0.05157750844955444, 'reward_total_composite_std': 0.0837271586060524, 'epoch': 0.01}

  7%|▋         | 223/3300 [48:20<8:59:22, 10.52s/it]INFO 04-12 11:34:33 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 224/3300 [48:29<8:33:46, 10.02s/it]2026-04-12 11:34:41,882 | INFO | train_grpo_train | metrics_logged mode=train step=224

                                                    
{'loss': 0.1064, 'grad_norm': 12.024731636047363, 'learning_rate': 9.324242424242424e-06, 'num_tokens': 462771.0, 'completions/mean_length': 52.875, 'completions/min_length': 39.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.875, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.5638595223426819, 'rewards/meter/std': 0.43769386410713196, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.8443635702133179, 'rewards/lexical_plausibility/std': 0.16386772692203522, 'rewards/judge_quality/mean': 0.48875001072883606, 'rewards/judge_quality/std': 0.3665841519832611, 'rewards/repeat_penalty/mean': 0.9855141639709473, 'rewards/repeat_penalty/std': 0.01785101927816868, 'rewards/near_duplicate_penalty/mean': 0.9855141639709473, 'rewards/near_duplicate_penalty/std': 0.01785101927816868, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2302250862121582, 'rewards/total_composite/std': 0.2972378134727478, 'reward': 0.2302250862121582, 'reward_std': 0.2972378134727478, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20806123316287994, 'sampling/sampling_logp_difference/max': 1.2408835887908936, 'sampling/importance_sampling_ratio/min': 0.33881494402885437, 'sampling/importance_sampling_ratio/mean': 1.0378979444503784, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6758073568344116, 'clip_ratio/low_mean': 0.12764441035687923, 'clip_ratio/low_min': 0.12764441035687923, 'clip_ratio/high_mean': 0.08768994174897671, 'clip_ratio/high_max': 0.08768994174897671, 'clip_ratio/region_mean': 0.21533435210585594, 'reward_total_mean': 0.2302250862121582, 'reward_meter_mean': 0.5638595223426819, 'reward_meter_std': 0.43769386410713196, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.8443635702133179, 'reward_lexical_plausibility_std': 0.16386772692203522, 'reward_repeat_penalty_mean': 0.9855141639709473, 'reward_repeat_penalty_std': 0.01785101927816868, 'reward_near_duplicate_penalty_mean': 0.9855141639709473, 'reward_near_duplicate_penalty_std': 0.01785101927816868, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.48875001072883606, 'reward_judge_quality_std': 0.3665841519832611, 'reward_total_composite_mean': 0.2302250862121582, 'reward_total_composite_std': 0.2972378134727478, 'epoch': 0.01}

  7%|▋         | 224/3300 [48:29<8:33:46, 10.02s/it]INFO 04-12 11:34:42 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 225/3300 [48:42<9:09:24, 10.72s/it]2026-04-12 11:34:54,231 | INFO | train_grpo_train | metrics_logged mode=train step=225

                                                    
{'loss': 0.2284, 'grad_norm': 8.010241508483887, 'learning_rate': 9.321212121212122e-06, 'num_tokens': 465156.0, 'completions/mean_length': 115.125, 'completions/min_length': 66.0, 'completions/max_length': 328.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 115.125, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 328.0, 'rewards/meter/mean': 0.23720678687095642, 'rewards/meter/std': 0.33528557419776917, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.8585575819015503, 'rewards/lexical_plausibility/std': 0.23015759885311127, 'rewards/judge_quality/mean': 0.20000000298023224, 'rewards/judge_quality/std': 0.13093073666095734, 'rewards/repeat_penalty/mean': 0.9812591075897217, 'rewards/repeat_penalty/std': 0.043902330100536346, 'rewards/near_duplicate_penalty/mean': 0.9906070232391357, 'rewards/near_duplicate_penalty/std': 0.017759235575795174, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.990138053894043, 'rewards/distinct_2/std': 0.027893755584955215, 'rewards/total_composite/mean': 0.03885792940855026, 'rewards/total_composite/std': 0.04952066391706467, 'reward': 0.03885792940855026, 'reward_std': 0.04952066019177437, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2218947559595108, 'sampling/sampling_logp_difference/max': 1.4545669555664062, 'sampling/importance_sampling_ratio/min': 0.23350146412849426, 'sampling/importance_sampling_ratio/mean': 1.040709137916565, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.6402252316474915, 'clip_ratio/low_mean': 0.09260035492479801, 'clip_ratio/low_min': 0.09260035492479801, 'clip_ratio/high_mean': 0.08005536161363125, 'clip_ratio/high_max': 0.08005536161363125, 'clip_ratio/region_mean': 0.17265571653842926, 'reward_total_mean': 0.03885792940855026, 'reward_meter_mean': 0.23720678687095642, 'reward_meter_std': 0.33528557419776917, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.8585575819015503, 'reward_lexical_plausibility_std': 0.23015759885311127, 'reward_repeat_penalty_mean': 0.9812591075897217, 'reward_repeat_penalty_std': 0.043902330100536346, 'reward_near_duplicate_penalty_mean': 0.9906070232391357, 'reward_near_duplicate_penalty_std': 0.017759235575795174, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.990138053894043, 'reward_distinct_2_std': 0.027893755584955215, 'reward_judge_quality_mean': 0.20000000298023224, 'reward_judge_quality_std': 0.13093073666095734, 'reward_total_composite_mean': 0.03885792940855026, 'reward_total_composite_std': 0.04952066391706467, 'epoch': 0.01}

  7%|▋         | 225/3300 [48:42<9:09:24, 10.72s/it]INFO 04-12 11:34:54 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 226/3300 [48:52<9:03:51, 10.62s/it]2026-04-12 11:35:04,602 | INFO | train_grpo_train | metrics_logged mode=train step=226

                                                    
{'loss': 0.0539, 'grad_norm': 8.86965560913086, 'learning_rate': 9.318181818181819e-06, 'num_tokens': 467685.0, 'completions/mean_length': 113.125, 'completions/min_length': 85.0, 'completions/max_length': 141.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 113.125, 'completions/min_terminated_length': 85.0, 'completions/max_terminated_length': 141.0, 'rewards/meter/mean': 0.5976791977882385, 'rewards/meter/std': 0.35980021953582764, 'rewards/count_adherence/mean': 0.9166666269302368, 'rewards/count_adherence/std': 0.1259881556034088, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9270759224891663, 'rewards/lexical_plausibility/std': 0.13093377649784088, 'rewards/judge_quality/mean': 0.17500001192092896, 'rewards/judge_quality/std': 0.08864051848649979, 'rewards/repeat_penalty/mean': 0.985623836517334, 'rewards/repeat_penalty/std': 0.02347756177186966, 'rewards/near_duplicate_penalty/mean': 0.9952421188354492, 'rewards/near_duplicate_penalty/std': 0.00500070583075285, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9902669787406921, 'rewards/distinct_2/std': 0.019538307562470436, 'rewards/total_composite/mean': 0.10001139342784882, 'rewards/total_composite/std': 0.11458198726177216, 'reward': 0.10001139342784882, 'reward_std': 0.11458198726177216, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2474474012851715, 'sampling/sampling_logp_difference/max': 1.7313833236694336, 'sampling/importance_sampling_ratio/min': 0.17703934013843536, 'sampling/importance_sampling_ratio/mean': 1.0650231838226318, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 4.06893390417099, 'clip_ratio/low_mean': 0.11227107793092728, 'clip_ratio/low_min': 0.11227107793092728, 'clip_ratio/high_mean': 0.08465197682380676, 'clip_ratio/high_max': 0.08465197682380676, 'clip_ratio/region_mean': 0.19692305475473404, 'reward_total_mean': 0.10001139342784882, 'reward_meter_mean': 0.5976791977882385, 'reward_meter_std': 0.35980021953582764, 'reward_count_adherence_mean': 0.9166666269302368, 'reward_count_adherence_std': 0.1259881556034088, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9270759224891663, 'reward_lexical_plausibility_std': 0.13093377649784088, 'reward_repeat_penalty_mean': 0.985623836517334, 'reward_repeat_penalty_std': 0.02347756177186966, 'reward_near_duplicate_penalty_mean': 0.9952421188354492, 'reward_near_duplicate_penalty_std': 0.00500070583075285, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9902669787406921, 'reward_distinct_2_std': 0.019538307562470436, 'reward_judge_quality_mean': 0.17500001192092896, 'reward_judge_quality_std': 0.08864051848649979, 'reward_total_composite_mean': 0.10001139342784882, 'reward_total_composite_std': 0.11458198726177216, 'epoch': 0.01}

  7%|▋         | 226/3300 [48:52<9:03:51, 10.62s/it]INFO 04-12 11:35:04 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 227/3300 [49:01<8:38:01, 10.11s/it]2026-04-12 11:35:13,548 | INFO | train_grpo_train | metrics_logged mode=train step=227

                                                    
{'loss': 0.1287, 'grad_norm': 16.40882682800293, 'learning_rate': 9.315151515151516e-06, 'num_tokens': 469261.0, 'completions/mean_length': 39.0, 'completions/min_length': 30.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.0, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.6159238815307617, 'rewards/meter/std': 0.41157564520835876, '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/lexical_plausibility/mean': 0.9756944179534912, 'rewards/lexical_plausibility/std': 0.058338064700365067, 'rewards/judge_quality/mean': 0.5387500524520874, 'rewards/judge_quality/std': 0.3228638768196106, 'rewards/repeat_penalty/mean': 0.9763433933258057, 'rewards/repeat_penalty/std': 0.044303566217422485, 'rewards/near_duplicate_penalty/mean': 0.9861366748809814, 'rewards/near_duplicate_penalty/std': 0.027932431548833847, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9898785352706909, 'rewards/distinct_2/std': 0.028627805411815643, 'rewards/total_composite/mean': 0.3676019310951233, 'rewards/total_composite/std': 0.3981827199459076, 'reward': 0.3676019310951233, 'reward_std': 0.3981827199459076, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2086174041032791, 'sampling/sampling_logp_difference/max': 1.6819677352905273, 'sampling/importance_sampling_ratio/min': 0.18600760400295258, 'sampling/importance_sampling_ratio/mean': 1.0399211645126343, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1516895294189453, 'clip_ratio/low_mean': 0.1118665412068367, 'clip_ratio/low_min': 0.1118665412068367, 'clip_ratio/high_mean': 0.0500331218354404, 'clip_ratio/high_max': 0.0500331218354404, 'clip_ratio/region_mean': 0.1618996630422771, 'reward_total_mean': 0.3676019310951233, 'reward_meter_mean': 0.6159238815307617, 'reward_meter_std': 0.41157564520835876, '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_lexical_plausibility_mean': 0.9756944179534912, 'reward_lexical_plausibility_std': 0.058338064700365067, 'reward_repeat_penalty_mean': 0.9763433933258057, 'reward_repeat_penalty_std': 0.044303566217422485, 'reward_near_duplicate_penalty_mean': 0.9861366748809814, 'reward_near_duplicate_penalty_std': 0.027932431548833847, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9898785352706909, 'reward_distinct_2_std': 0.028627805411815643, 'reward_judge_quality_mean': 0.5387500524520874, 'reward_judge_quality_std': 0.3228638768196106, 'reward_total_composite_mean': 0.3676019310951233, 'reward_total_composite_std': 0.3981827199459076, 'epoch': 0.01}

  7%|▋         | 227/3300 [49:01<8:38:01, 10.11s/it]INFO 04-12 11:35:13 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 228/3300 [49:10<8:25:18,  9.87s/it]2026-04-12 11:35:22,845 | INFO | train_grpo_train | metrics_logged mode=train step=228

                                                    
{'loss': -0.0375, 'grad_norm': 28.835786819458008, 'learning_rate': 9.312121212121212e-06, 'num_tokens': 470879.0, 'completions/mean_length': 29.25, 'completions/min_length': 23.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.25, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.28039538860321045, 'rewards/meter/std': 0.25342094898223877, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9708828330039978, 'rewards/lexical_plausibility/std': 0.06114698573946953, 'rewards/judge_quality/mean': 0.4337500035762787, 'rewards/judge_quality/std': 0.19956651329994202, 'rewards/repeat_penalty/mean': 0.9233226180076599, 'rewards/repeat_penalty/std': 0.17184801399707794, 'rewards/near_duplicate_penalty/mean': 0.9658033847808838, 'rewards/near_duplicate_penalty/std': 0.055021733045578, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9743589758872986, 'rewards/distinct_2/std': 0.07252377271652222, 'rewards/total_composite/mean': 0.11529844254255295, 'rewards/total_composite/std': 0.09825806319713593, 'reward': 0.11529844254255295, 'reward_std': 0.09825806319713593, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20324234664440155, 'sampling/sampling_logp_difference/max': 2.6962225437164307, 'sampling/importance_sampling_ratio/min': 0.06745985895395279, 'sampling/importance_sampling_ratio/mean': 1.0252653360366821, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.840165987610817, 'clip_ratio/low_mean': 0.10353380255401134, 'clip_ratio/low_min': 0.10353380255401134, 'clip_ratio/high_mean': 0.08810030948370695, 'clip_ratio/high_max': 0.08810030948370695, 'clip_ratio/region_mean': 0.1916341120377183, 'reward_total_mean': 0.11529844254255295, 'reward_meter_mean': 0.28039538860321045, 'reward_meter_std': 0.25342094898223877, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9708828330039978, 'reward_lexical_plausibility_std': 0.06114698573946953, 'reward_repeat_penalty_mean': 0.9233226180076599, 'reward_repeat_penalty_std': 0.17184801399707794, 'reward_near_duplicate_penalty_mean': 0.9658033847808838, 'reward_near_duplicate_penalty_std': 0.055021733045578, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9743589758872986, 'reward_distinct_2_std': 0.07252377271652222, 'reward_judge_quality_mean': 0.4337500035762787, 'reward_judge_quality_std': 0.19956651329994202, 'reward_total_composite_mean': 0.11529844254255295, 'reward_total_composite_std': 0.09825806319713593, 'epoch': 0.01}

  7%|▋         | 228/3300 [49:10<8:25:18,  9.87s/it]INFO 04-12 11:35:23 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 229/3300 [49:19<8:07:06,  9.52s/it]2026-04-12 11:35:31,540 | INFO | train_grpo_train | metrics_logged mode=train step=229

                                                    
{'loss': 0.0833, 'grad_norm': 19.24261474609375, 'learning_rate': 9.30909090909091e-06, 'num_tokens': 472231.0, 'completions/mean_length': 24.0, 'completions/min_length': 18.0, 'completions/max_length': 31.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.0, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.48322755098342896, 'rewards/meter/std': 0.4733370840549469, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9752674102783203, 'rewards/lexical_plausibility/std': 0.06995441019535065, 'rewards/judge_quality/mean': 0.5087499618530273, 'rewards/judge_quality/std': 0.2861786484718323, 'rewards/repeat_penalty/mean': 0.7130681872367859, 'rewards/repeat_penalty/std': 0.10445895045995712, 'rewards/near_duplicate_penalty/mean': 0.9507575631141663, 'rewards/near_duplicate_penalty/std': 0.1392786055803299, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.20232832431793213, 'rewards/total_composite/std': 0.1917257457971573, 'reward': 0.20232832431793213, 'reward_std': 0.1917257308959961, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21530920267105103, 'sampling/sampling_logp_difference/max': 1.189661979675293, 'sampling/importance_sampling_ratio/min': 0.30432412028312683, 'sampling/importance_sampling_ratio/mean': 1.0869141817092896, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9509296268224716, 'clip_ratio/low_mean': 0.11066515184938908, 'clip_ratio/low_min': 0.11066515184938908, 'clip_ratio/high_mean': 0.08464077021926641, 'clip_ratio/high_max': 0.08464077021926641, 'clip_ratio/region_mean': 0.1953059220686555, 'reward_total_mean': 0.20232832431793213, 'reward_meter_mean': 0.48322755098342896, 'reward_meter_std': 0.4733370840549469, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9752674102783203, 'reward_lexical_plausibility_std': 0.06995441019535065, 'reward_repeat_penalty_mean': 0.7130681872367859, 'reward_repeat_penalty_std': 0.10445895045995712, 'reward_near_duplicate_penalty_mean': 0.9507575631141663, 'reward_near_duplicate_penalty_std': 0.1392786055803299, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5087499618530273, 'reward_judge_quality_std': 0.2861786484718323, 'reward_total_composite_mean': 0.20232832431793213, 'reward_total_composite_std': 0.1917257457971573, 'epoch': 0.01}

  7%|▋         | 229/3300 [49:19<8:07:06,  9.52s/it]INFO 04-12 11:35:31 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 230/3300 [49:29<8:15:11,  9.68s/it]2026-04-12 11:35:41,594 | INFO | train_grpo_train | metrics_logged mode=train step=230

                                                    
{'loss': 0.0721, 'grad_norm': 8.41430950164795, 'learning_rate': 9.306060606060608e-06, 'num_tokens': 474661.0, 'completions/mean_length': 102.75, 'completions/min_length': 80.0, 'completions/max_length': 141.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 102.75, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 141.0, 'rewards/meter/mean': 0.41917771100997925, 'rewards/meter/std': 0.31770917773246765, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9525266289710999, 'rewards/lexical_plausibility/std': 0.10706377774477005, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.9959983229637146, 'rewards/repeat_penalty/std': 0.004256305750459433, 'rewards/near_duplicate_penalty/mean': 0.9959983229637146, 'rewards/near_duplicate_penalty/std': 0.004256305750459433, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1410641372203827, 'rewards/total_composite/std': 0.13569699227809906, 'reward': 0.1410641372203827, 'reward_std': 0.13569699227809906, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24845126271247864, 'sampling/sampling_logp_difference/max': 1.806035041809082, 'sampling/importance_sampling_ratio/min': 0.16430431604385376, 'sampling/importance_sampling_ratio/mean': 1.0583014488220215, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.6665093302726746, 'clip_ratio/low_mean': 0.09639752097427845, 'clip_ratio/low_min': 0.09639752097427845, 'clip_ratio/high_mean': 0.08284858800470829, 'clip_ratio/high_max': 0.08284858800470829, 'clip_ratio/region_mean': 0.17924610897898674, 'reward_total_mean': 0.1410641372203827, 'reward_meter_mean': 0.41917771100997925, 'reward_meter_std': 0.31770917773246765, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9525266289710999, 'reward_lexical_plausibility_std': 0.10706377774477005, 'reward_repeat_penalty_mean': 0.9959983229637146, 'reward_repeat_penalty_std': 0.004256305750459433, 'reward_near_duplicate_penalty_mean': 0.9959983229637146, 'reward_near_duplicate_penalty_std': 0.004256305750459433, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.1410641372203827, 'reward_total_composite_std': 0.13569699227809906, 'epoch': 0.01}

  7%|▋         | 230/3300 [49:29<8:15:11,  9.68s/it]INFO 04-12 11:35:41 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 231/3300 [49:38<8:06:53,  9.52s/it]2026-04-12 11:35:50,742 | INFO | train_grpo_train | metrics_logged mode=train step=231

                                                    
{'loss': 0.016, 'grad_norm': 10.604702949523926, 'learning_rate': 9.303030303030303e-06, 'num_tokens': 476493.0, 'completions/mean_length': 73.0, 'completions/min_length': 60.0, 'completions/max_length': 95.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.0, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 95.0, 'rewards/meter/mean': 0.39129316806793213, 'rewards/meter/std': 0.26790088415145874, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9722594022750854, 'rewards/lexical_plausibility/std': 0.07774464040994644, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.16035674512386322, 'rewards/repeat_penalty/mean': 0.9915815591812134, 'rewards/repeat_penalty/std': 0.006199967116117477, 'rewards/near_duplicate_penalty/mean': 0.9915815591812134, 'rewards/near_duplicate_penalty/std': 0.006199967116117477, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.11387395858764648, 'rewards/total_composite/std': 0.11799918860197067, 'reward': 0.11387395858764648, 'reward_std': 0.11799918860197067, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2647823393344879, 'sampling/sampling_logp_difference/max': 1.5168256759643555, 'sampling/importance_sampling_ratio/min': 0.21940726041793823, 'sampling/importance_sampling_ratio/mean': 1.0522475242614746, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.499101161956787, 'clip_ratio/low_mean': 0.12810280360281467, 'clip_ratio/low_min': 0.12810280360281467, 'clip_ratio/high_mean': 0.0863079372793436, 'clip_ratio/high_max': 0.0863079372793436, 'clip_ratio/region_mean': 0.21441074088215828, 'reward_total_mean': 0.11387395858764648, 'reward_meter_mean': 0.39129316806793213, 'reward_meter_std': 0.26790088415145874, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9722594022750854, 'reward_lexical_plausibility_std': 0.07774464040994644, 'reward_repeat_penalty_mean': 0.9915815591812134, 'reward_repeat_penalty_std': 0.006199967116117477, 'reward_near_duplicate_penalty_mean': 0.9915815591812134, 'reward_near_duplicate_penalty_std': 0.006199967116117477, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.16035674512386322, 'reward_total_composite_mean': 0.11387395858764648, 'reward_total_composite_std': 0.11799918860197067, 'epoch': 0.01}

  7%|▋         | 231/3300 [49:38<8:06:53,  9.52s/it]INFO 04-12 11:35:51 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 232/3300 [49:53<9:28:32, 11.12s/it]2026-04-12 11:36:05,593 | INFO | train_grpo_train | metrics_logged mode=train step=232

                                                    
{'loss': -0.0352, 'grad_norm': 5.134120941162109, 'learning_rate': 9.3e-06, 'num_tokens': 478782.0, 'completions/mean_length': 165.125, 'completions/min_length': 99.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 115.5714340209961, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 130.0, 'rewards/meter/mean': 0.5854902267456055, 'rewards/meter/std': 0.2812030017375946, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.2357022762298584, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.883344829082489, 'rewards/lexical_plausibility/std': 0.1693534404039383, 'rewards/judge_quality/mean': 0.21250000596046448, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.9939590692520142, 'rewards/repeat_penalty/std': 0.009112166240811348, 'rewards/near_duplicate_penalty/mean': 0.9969067573547363, 'rewards/near_duplicate_penalty/std': 0.003100398927927017, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9970414042472839, 'rewards/distinct_2/std': 0.008368130773305893, 'rewards/total_composite/mean': 0.12124696373939514, 'rewards/total_composite/std': 0.1220458373427391, 'reward': 0.12124696373939514, 'reward_std': 0.1220458373427391, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24998222291469574, 'sampling/sampling_logp_difference/max': 1.5013141632080078, 'sampling/importance_sampling_ratio/min': 0.22283713519573212, 'sampling/importance_sampling_ratio/mean': 1.0617833137512207, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.9537555873394012, 'clip_ratio/low_mean': 0.052912636660039425, 'clip_ratio/low_min': 0.052912636660039425, 'clip_ratio/high_mean': 0.11284627206623554, 'clip_ratio/high_max': 0.11284627206623554, 'clip_ratio/region_mean': 0.16575890872627497, 'reward_total_mean': 0.12124696373939514, 'reward_meter_mean': 0.5854902267456055, 'reward_meter_std': 0.2812030017375946, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.2357022762298584, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.883344829082489, 'reward_lexical_plausibility_std': 0.1693534404039383, 'reward_repeat_penalty_mean': 0.9939590692520142, 'reward_repeat_penalty_std': 0.009112166240811348, 'reward_near_duplicate_penalty_mean': 0.9969067573547363, 'reward_near_duplicate_penalty_std': 0.003100398927927017, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9970414042472839, 'reward_distinct_2_std': 0.008368130773305893, 'reward_judge_quality_mean': 0.21250000596046448, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.12124696373939514, 'reward_total_composite_std': 0.1220458373427391, 'epoch': 0.01}

  7%|▋         | 232/3300 [49:53<9:28:32, 11.12s/it]INFO 04-12 11:36:05 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 233/3300 [50:08<10:32:23, 12.37s/it]2026-04-12 11:36:20,888 | INFO | train_grpo_train | metrics_logged mode=train step=233

                                                     
{'loss': -0.0257, 'grad_norm': 4.961449146270752, 'learning_rate': 9.296969696969698e-06, 'num_tokens': 481089.0, 'completions/mean_length': 176.375, 'completions/min_length': 95.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 128.42857360839844, 'completions/min_terminated_length': 95.0, 'completions/max_terminated_length': 169.0, 'rewards/meter/mean': 0.49392178654670715, 'rewards/meter/std': 0.4410583972930908, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.8732752799987793, 'rewards/lexical_plausibility/std': 0.2317531555891037, 'rewards/judge_quality/mean': 0.19999998807907104, 'rewards/judge_quality/std': 0.13093073666095734, 'rewards/repeat_penalty/mean': 0.9971146583557129, 'rewards/repeat_penalty/std': 0.0030275366734713316, 'rewards/near_duplicate_penalty/mean': 0.9971146583557129, 'rewards/near_duplicate_penalty/std': 0.0030275366734713316, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.12325656414031982, 'rewards/total_composite/std': 0.1587062031030655, 'reward': 0.12325656414031982, 'reward_std': 0.1587061882019043, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24637077748775482, 'sampling/sampling_logp_difference/max': 1.6192541122436523, 'sampling/importance_sampling_ratio/min': 0.19804635643959045, 'sampling/importance_sampling_ratio/mean': 1.0744915008544922, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.3490015864372253, 'clip_ratio/low_mean': 0.062875647097826, 'clip_ratio/low_min': 0.062875647097826, 'clip_ratio/high_mean': 0.09082281030714512, 'clip_ratio/high_max': 0.09082281030714512, 'clip_ratio/region_mean': 0.15369845740497112, 'reward_total_mean': 0.12325656414031982, 'reward_meter_mean': 0.49392178654670715, 'reward_meter_std': 0.4410583972930908, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.09258200973272324, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.8732752799987793, 'reward_lexical_plausibility_std': 0.2317531555891037, 'reward_repeat_penalty_mean': 0.9971146583557129, 'reward_repeat_penalty_std': 0.0030275366734713316, 'reward_near_duplicate_penalty_mean': 0.9971146583557129, 'reward_near_duplicate_penalty_std': 0.0030275366734713316, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.19999998807907104, 'reward_judge_quality_std': 0.13093073666095734, 'reward_total_composite_mean': 0.12325656414031982, 'reward_total_composite_std': 0.1587062031030655, 'epoch': 0.01}

  7%|▋         | 233/3300 [50:08<10:32:23, 12.37s/it]INFO 04-12 11:36:21 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 234/3300 [50:18<9:55:28, 11.65s/it] 2026-04-12 11:36:30,867 | INFO | train_grpo_train | metrics_logged mode=train step=234

                                                    
{'loss': 0.1393, 'grad_norm': 16.711183547973633, 'learning_rate': 9.293939393939395e-06, 'num_tokens': 482720.0, 'completions/mean_length': 39.875, 'completions/min_length': 33.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.875, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.8442812561988831, 'rewards/meter/std': 0.22092205286026, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.1927248239517212, 'rewards/repeat_penalty/mean': 0.9821171164512634, 'rewards/repeat_penalty/std': 0.027723101899027824, 'rewards/near_duplicate_penalty/mean': 0.9911913871765137, 'rewards/near_duplicate_penalty/std': 0.010196942836046219, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9908424615859985, 'rewards/distinct_2/std': 0.025901345536112785, 'rewards/total_composite/mean': 0.2744794189929962, 'rewards/total_composite/std': 0.18476887047290802, 'reward': 0.2744794189929962, 'reward_std': 0.18476885557174683, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22389782965183258, 'sampling/sampling_logp_difference/max': 1.702122688293457, 'sampling/importance_sampling_ratio/min': 0.18229615688323975, 'sampling/importance_sampling_ratio/mean': 1.0627580881118774, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.865033984184265, 'clip_ratio/low_mean': 0.07573607377707958, 'clip_ratio/low_min': 0.07573607377707958, 'clip_ratio/high_mean': 0.08506254758685827, 'clip_ratio/high_max': 0.08506254758685827, 'clip_ratio/region_mean': 0.16079862136393785, 'reward_total_mean': 0.2744794189929962, 'reward_meter_mean': 0.8442812561988831, 'reward_meter_std': 0.22092205286026, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9821171164512634, 'reward_repeat_penalty_std': 0.027723101899027824, 'reward_near_duplicate_penalty_mean': 0.9911913871765137, 'reward_near_duplicate_penalty_std': 0.010196942836046219, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9908424615859985, 'reward_distinct_2_std': 0.025901345536112785, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.1927248239517212, 'reward_total_composite_mean': 0.2744794189929962, 'reward_total_composite_std': 0.18476887047290802, 'epoch': 0.01}

  7%|▋         | 234/3300 [50:18<9:55:28, 11.65s/it]INFO 04-12 11:36:31 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 235/3300 [50:30<10:01:06, 11.77s/it]2026-04-12 11:36:42,898 | INFO | train_grpo_train | metrics_logged mode=train step=235

                                                     
{'loss': 0.2869, 'grad_norm': 10.497565269470215, 'learning_rate': 9.29090909090909e-06, 'num_tokens': 484440.0, 'completions/mean_length': 68.0, 'completions/min_length': 30.0, 'completions/max_length': 261.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.0, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 261.0, 'rewards/meter/mean': 0.25676068663597107, 'rewards/meter/std': 0.31785857677459717, '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/lexical_plausibility/mean': 0.9099849462509155, 'rewards/lexical_plausibility/std': 0.2546009123325348, 'rewards/judge_quality/mean': 0.36249998211860657, 'rewards/judge_quality/std': 0.13562026619911194, 'rewards/repeat_penalty/mean': 0.9589519500732422, 'rewards/repeat_penalty/std': 0.046857405453920364, 'rewards/near_duplicate_penalty/mean': 0.9689867496490479, 'rewards/near_duplicate_penalty/std': 0.04364578798413277, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9898785352706909, 'rewards/distinct_2/std': 0.028627805411815643, 'rewards/total_composite/mean': 0.08154526352882385, 'rewards/total_composite/std': 0.13823302090168, 'reward': 0.08154526352882385, 'reward_std': 0.13823302090168, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2429894357919693, 'sampling/sampling_logp_difference/max': 1.8698291778564453, 'sampling/importance_sampling_ratio/min': 0.15414999425411224, 'sampling/importance_sampling_ratio/mean': 1.0274935960769653, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.7454141974449158, 'clip_ratio/low_mean': 0.13338296953588724, 'clip_ratio/low_min': 0.13338296953588724, 'clip_ratio/high_mean': 0.0690789483487606, 'clip_ratio/high_max': 0.0690789483487606, 'clip_ratio/region_mean': 0.20246191788464785, 'reward_total_mean': 0.08154526352882385, 'reward_meter_mean': 0.25676068663597107, 'reward_meter_std': 0.31785857677459717, '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_lexical_plausibility_mean': 0.9099849462509155, 'reward_lexical_plausibility_std': 0.2546009123325348, 'reward_repeat_penalty_mean': 0.9589519500732422, 'reward_repeat_penalty_std': 0.046857405453920364, 'reward_near_duplicate_penalty_mean': 0.9689867496490479, 'reward_near_duplicate_penalty_std': 0.04364578798413277, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9898785352706909, 'reward_distinct_2_std': 0.028627805411815643, 'reward_judge_quality_mean': 0.36249998211860657, 'reward_judge_quality_std': 0.13562026619911194, 'reward_total_composite_mean': 0.08154526352882385, 'reward_total_composite_std': 0.13823302090168, 'epoch': 0.01}

  7%|▋         | 235/3300 [50:30<10:01:06, 11.77s/it]INFO 04-12 11:36:43 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 236/3300 [50:45<10:45:39, 12.64s/it]2026-04-12 11:36:57,586 | INFO | train_grpo_train | metrics_logged mode=train step=236

                                                     
{'loss': -0.0726, 'grad_norm': 2.903174877166748, 'learning_rate': 9.28787878787879e-06, 'num_tokens': 487401.0, 'completions/mean_length': 389.125, 'completions/min_length': 186.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 266.25, 'completions/min_terminated_length': 186.0, 'completions/max_terminated_length': 466.0, 'rewards/meter/mean': 0.4403528571128845, 'rewards/meter/std': 0.3332734704017639, 'rewards/count_adherence/mean': 0.762499988079071, 'rewards/count_adherence/std': 0.2386719137430191, 'rewards/arabic_clean/mean': 0.375, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.6608182191848755, 'rewards/lexical_plausibility/std': 0.2512224018573761, 'rewards/judge_quality/mean': 0.125, 'rewards/judge_quality/std': 0.10350984334945679, 'rewards/repeat_penalty/mean': 0.9976162910461426, 'rewards/repeat_penalty/std': 0.001307821017690003, 'rewards/near_duplicate_penalty/mean': 0.9976162910461426, 'rewards/near_duplicate_penalty/std': 0.001307821017690003, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.05369948223233223, 'rewards/total_composite/std': 0.12017948925495148, 'reward': 0.05369948223233223, 'reward_std': 0.12017949670553207, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21965289115905762, 'sampling/sampling_logp_difference/max': 1.0888519287109375, 'sampling/importance_sampling_ratio/min': 0.33660271763801575, 'sampling/importance_sampling_ratio/mean': 1.0608978271484375, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.7886303663253784, 'clip_ratio/low_mean': 0.03072961326688528, 'clip_ratio/low_min': 0.03072961326688528, 'clip_ratio/high_mean': 0.05858132615685463, 'clip_ratio/high_max': 0.05858132615685463, 'clip_ratio/region_mean': 0.08931093942373991, 'reward_total_mean': 0.05369948223233223, 'reward_meter_mean': 0.4403528571128845, 'reward_meter_std': 0.3332734704017639, 'reward_count_adherence_mean': 0.762499988079071, 'reward_count_adherence_std': 0.2386719137430191, 'reward_arabic_clean_mean': 0.375, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.6608182191848755, 'reward_lexical_plausibility_std': 0.2512224018573761, 'reward_repeat_penalty_mean': 0.9976162910461426, 'reward_repeat_penalty_std': 0.001307821017690003, 'reward_near_duplicate_penalty_mean': 0.9976162910461426, 'reward_near_duplicate_penalty_std': 0.001307821017690003, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.125, 'reward_judge_quality_std': 0.10350984334945679, 'reward_total_composite_mean': 0.05369948223233223, 'reward_total_composite_std': 0.12017948925495148, 'epoch': 0.01}

  7%|▋         | 236/3300 [50:45<10:45:39, 12.64s/it]INFO 04-12 11:36:57 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 237/3300 [51:02<11:44:11, 13.79s/it]2026-04-12 11:37:14,073 | INFO | train_grpo_train | metrics_logged mode=train step=237

                                                     
{'loss': 0.0256, 'grad_norm': 7.149643421173096, 'learning_rate': 9.284848484848485e-06, 'num_tokens': 489016.0, 'completions/mean_length': 98.875, 'completions/min_length': 27.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 39.85714340209961, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.4038103222846985, 'rewards/meter/std': 0.3540455400943756, '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/lexical_plausibility/mean': 0.9055606722831726, 'rewards/lexical_plausibility/std': 0.21432743966579437, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.155264750123024, 'rewards/repeat_penalty/mean': 0.9937461614608765, 'rewards/repeat_penalty/std': 0.0076161217875778675, 'rewards/near_duplicate_penalty/mean': 0.9962989091873169, 'rewards/near_duplicate_penalty/std': 0.005241013132035732, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9974472522735596, 'rewards/distinct_2/std': 0.007220282219350338, 'rewards/total_composite/mean': 0.13621383905410767, 'rewards/total_composite/std': 0.1638379842042923, 'reward': 0.13621383905410767, 'reward_std': 0.1638379842042923, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2056620866060257, 'sampling/sampling_logp_difference/max': 1.485896348953247, 'sampling/importance_sampling_ratio/min': 0.22629940509796143, 'sampling/importance_sampling_ratio/mean': 1.0489250421524048, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8136246800422668, 'clip_ratio/low_mean': 0.08420171681791544, 'clip_ratio/low_min': 0.08420171681791544, 'clip_ratio/high_mean': 0.08618157543241978, 'clip_ratio/high_max': 0.08618157543241978, 'clip_ratio/region_mean': 0.17038329225033522, 'reward_total_mean': 0.13621383905410767, 'reward_meter_mean': 0.4038103222846985, 'reward_meter_std': 0.3540455400943756, '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_lexical_plausibility_mean': 0.9055606722831726, 'reward_lexical_plausibility_std': 0.21432743966579437, 'reward_repeat_penalty_mean': 0.9937461614608765, 'reward_repeat_penalty_std': 0.0076161217875778675, 'reward_near_duplicate_penalty_mean': 0.9962989091873169, 'reward_near_duplicate_penalty_std': 0.005241013132035732, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9974472522735596, 'reward_distinct_2_std': 0.007220282219350338, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.155264750123024, 'reward_total_composite_mean': 0.13621383905410767, 'reward_total_composite_std': 0.1638379842042923, 'epoch': 0.01}

  7%|▋         | 237/3300 [51:02<11:44:11, 13.79s/it]INFO 04-12 11:37:14 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 238/3300 [51:13<11:08:04, 13.09s/it]2026-04-12 11:37:25,524 | INFO | train_grpo_train | metrics_logged mode=train step=238

                                                     
{'loss': 0.0505, 'grad_norm': 16.314659118652344, 'learning_rate': 9.281818181818183e-06, 'num_tokens': 490652.0, 'completions/mean_length': 52.5, 'completions/min_length': 44.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.5, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.6325927376747131, 'rewards/meter/std': 0.2964901924133301, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9518710374832153, 'rewards/lexical_plausibility/std': 0.09042418748140335, 'rewards/judge_quality/mean': 0.5074999928474426, 'rewards/judge_quality/std': 0.2516658902168274, 'rewards/repeat_penalty/mean': 0.9930394887924194, 'rewards/repeat_penalty/std': 0.010083186440169811, 'rewards/near_duplicate_penalty/mean': 0.9930394887924194, 'rewards/near_duplicate_penalty/std': 0.010083186440169811, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.32204437255859375, 'rewards/total_composite/std': 0.22403883934020996, 'reward': 0.32204437255859375, 'reward_std': 0.22403883934020996, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22436624765396118, 'sampling/sampling_logp_difference/max': 2.238736152648926, 'sampling/importance_sampling_ratio/min': 0.10659313201904297, 'sampling/importance_sampling_ratio/mean': 0.9726545214653015, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4239566251635551, 'clip_ratio/low_mean': 0.08723804075270891, 'clip_ratio/low_min': 0.08723804075270891, 'clip_ratio/high_mean': 0.09302003309130669, 'clip_ratio/high_max': 0.09302003309130669, 'clip_ratio/region_mean': 0.1802580738440156, 'reward_total_mean': 0.32204437255859375, 'reward_meter_mean': 0.6325927376747131, 'reward_meter_std': 0.2964901924133301, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9518710374832153, 'reward_lexical_plausibility_std': 0.09042418748140335, 'reward_repeat_penalty_mean': 0.9930394887924194, 'reward_repeat_penalty_std': 0.010083186440169811, 'reward_near_duplicate_penalty_mean': 0.9930394887924194, 'reward_near_duplicate_penalty_std': 0.010083186440169811, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5074999928474426, 'reward_judge_quality_std': 0.2516658902168274, 'reward_total_composite_mean': 0.32204437255859375, 'reward_total_composite_std': 0.22403883934020996, 'epoch': 0.01}

  7%|▋         | 238/3300 [51:13<11:08:04, 13.09s/it]INFO 04-12 11:37:25 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 239/3300 [51:22<10:12:28, 12.01s/it]2026-04-12 11:37:34,988 | INFO | train_grpo_train | metrics_logged mode=train step=239

                                                     
{'loss': -0.024, 'grad_norm': 14.57002067565918, 'learning_rate': 9.27878787878788e-06, 'num_tokens': 491955.0, 'completions/mean_length': 16.875, 'completions/min_length': 16.0, 'completions/max_length': 18.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 16.875, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 18.0, 'rewards/meter/mean': 0.7849218845367432, 'rewards/meter/std': 0.26776716113090515, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.76500004529953, 'rewards/judge_quality/std': 0.28824594616889954, 'rewards/repeat_penalty/mean': 0.6820960640907288, 'rewards/repeat_penalty/std': 0.019698383286595345, 'rewards/near_duplicate_penalty/mean': 0.909461498260498, 'rewards/near_duplicate_penalty/std': 0.02626452036201954, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5813707709312439, 'rewards/total_composite/std': 0.3000923693180084, 'reward': 0.5813707709312439, 'reward_std': 0.3000923991203308, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07100889086723328, 'sampling/sampling_logp_difference/max': 0.8280153274536133, 'sampling/importance_sampling_ratio/min': 0.4369155764579773, 'sampling/importance_sampling_ratio/mean': 1.0254554748535156, 'sampling/importance_sampling_ratio/max': 1.6485719680786133, 'entropy': 0.497576754540205, 'clip_ratio/low_mean': 0.02083333395421505, 'clip_ratio/low_min': 0.02083333395421505, 'clip_ratio/high_mean': 0.014705882407724857, 'clip_ratio/high_max': 0.014705882407724857, 'clip_ratio/region_mean': 0.03553921636193991, 'reward_total_mean': 0.5813707709312439, 'reward_meter_mean': 0.7849218845367432, 'reward_meter_std': 0.26776716113090515, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6820960640907288, 'reward_repeat_penalty_std': 0.019698383286595345, 'reward_near_duplicate_penalty_mean': 0.909461498260498, 'reward_near_duplicate_penalty_std': 0.02626452036201954, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.76500004529953, 'reward_judge_quality_std': 0.28824594616889954, 'reward_total_composite_mean': 0.5813707709312439, 'reward_total_composite_std': 0.3000923693180084, 'epoch': 0.01}

  7%|▋         | 239/3300 [51:22<10:12:28, 12.01s/it]INFO 04-12 11:37:35 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 240/3300 [51:32<9:39:53, 11.37s/it] 2026-04-12 11:37:44,896 | INFO | train_grpo_train | metrics_logged mode=train step=240

                                                    
{'loss': 0.0513, 'grad_norm': 20.61700439453125, 'learning_rate': 9.275757575757577e-06, 'num_tokens': 493473.0, 'completions/mean_length': 21.75, 'completions/min_length': 16.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.75, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.6308422684669495, 'rewards/meter/std': 0.48971638083457947, '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/lexical_plausibility/mean': 0.9771242141723633, 'rewards/lexical_plausibility/std': 0.0647025778889656, 'rewards/judge_quality/mean': 0.42124998569488525, 'rewards/judge_quality/std': 0.24032345414161682, 'rewards/repeat_penalty/mean': 0.78125, 'rewards/repeat_penalty/std': 0.0883883461356163, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3269261121749878, 'rewards/total_composite/std': 0.30889177322387695, 'reward': 0.3269261121749878, 'reward_std': 0.30889180302619934, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2429298758506775, 'sampling/sampling_logp_difference/max': 1.5405826568603516, 'sampling/importance_sampling_ratio/min': 0.21425624191761017, 'sampling/importance_sampling_ratio/mean': 1.0725314617156982, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.5414575040340424, 'clip_ratio/low_mean': 0.04205827065743506, 'clip_ratio/low_min': 0.04205827065743506, 'clip_ratio/high_mean': 0.12063401564955711, 'clip_ratio/high_max': 0.12063401564955711, 'clip_ratio/region_mean': 0.16269228630699217, 'reward_total_mean': 0.3269261121749878, 'reward_meter_mean': 0.6308422684669495, 'reward_meter_std': 0.48971638083457947, '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_lexical_plausibility_mean': 0.9771242141723633, 'reward_lexical_plausibility_std': 0.0647025778889656, 'reward_repeat_penalty_mean': 0.78125, 'reward_repeat_penalty_std': 0.0883883461356163, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.42124998569488525, 'reward_judge_quality_std': 0.24032345414161682, 'reward_total_composite_mean': 0.3269261121749878, 'reward_total_composite_std': 0.30889177322387695, 'epoch': 0.01}

  7%|▋         | 240/3300 [51:32<9:39:53, 11.37s/it]INFO 04-12 11:37:45 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 241/3300 [51:47<10:24:02, 12.24s/it]2026-04-12 11:37:59,145 | INFO | train_grpo_train | metrics_logged mode=train step=241

                                                     
{'loss': -0.047, 'grad_norm': 3.39906644821167, 'learning_rate': 9.272727272727273e-06, 'num_tokens': 495972.0, 'completions/mean_length': 383.375, 'completions/min_length': 164.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 254.75, 'completions/min_terminated_length': 164.0, 'completions/max_terminated_length': 429.0, 'rewards/meter/mean': 0.23733344674110413, 'rewards/meter/std': 0.3067074418067932, 'rewards/count_adherence/mean': 0.828125, 'rewards/count_adherence/std': 0.22097088396549225, 'rewards/arabic_clean/mean': 0.375, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.5411629676818848, 'rewards/lexical_plausibility/std': 0.2671348750591278, 'rewards/judge_quality/mean': 0.125, 'rewards/judge_quality/std': 0.10350984334945679, 'rewards/repeat_penalty/mean': 0.9860943555831909, 'rewards/repeat_penalty/std': 0.01868544891476631, 'rewards/near_duplicate_penalty/mean': 0.9923439025878906, 'rewards/near_duplicate_penalty/std': 0.006741866935044527, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9936693906784058, 'rewards/distinct_2/std': 0.015382092446088791, 'rewards/total_composite/mean': 0.02013734169304371, 'rewards/total_composite/std': 0.050523772835731506, 'reward': 0.02013734169304371, 'reward_std': 0.05052376911044121, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22711551189422607, 'sampling/sampling_logp_difference/max': 1.206812858581543, 'sampling/importance_sampling_ratio/min': 0.29914918541908264, 'sampling/importance_sampling_ratio/mean': 1.0702348947525024, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.358061909675598, 'clip_ratio/low_mean': 0.059528362937271595, 'clip_ratio/low_min': 0.059528362937271595, 'clip_ratio/high_mean': 0.02049180306494236, 'clip_ratio/high_max': 0.02049180306494236, 'clip_ratio/region_mean': 0.08002016600221395, 'reward_total_mean': 0.02013734169304371, 'reward_meter_mean': 0.23733344674110413, 'reward_meter_std': 0.3067074418067932, 'reward_count_adherence_mean': 0.828125, 'reward_count_adherence_std': 0.22097088396549225, 'reward_arabic_clean_mean': 0.375, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.5411629676818848, 'reward_lexical_plausibility_std': 0.2671348750591278, 'reward_repeat_penalty_mean': 0.9860943555831909, 'reward_repeat_penalty_std': 0.01868544891476631, 'reward_near_duplicate_penalty_mean': 0.9923439025878906, 'reward_near_duplicate_penalty_std': 0.006741866935044527, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9936693906784058, 'reward_distinct_2_std': 0.015382092446088791, 'reward_judge_quality_mean': 0.125, 'reward_judge_quality_std': 0.10350984334945679, 'reward_total_composite_mean': 0.02013734169304371, 'reward_total_composite_std': 0.050523772835731506, 'epoch': 0.01}

  7%|▋         | 241/3300 [51:47<10:24:02, 12.24s/it]INFO 04-12 11:37:59 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 242/3300 [52:01<10:58:46, 12.93s/it]2026-04-12 11:38:13,678 | INFO | train_grpo_train | metrics_logged mode=train step=242

                                                     
{'loss': -0.0496, 'grad_norm': 3.1483123302459717, 'learning_rate': 9.26969696969697e-06, 'num_tokens': 497703.0, 'completions/mean_length': 176.375, 'completions/min_length': 46.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 64.5, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.4226337671279907, 'rewards/meter/std': 0.3670276403427124, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.8753880858421326, 'rewards/lexical_plausibility/std': 0.2329125553369522, 'rewards/judge_quality/mean': 0.2749999761581421, 'rewards/judge_quality/std': 0.16690459847450256, 'rewards/repeat_penalty/mean': 0.9990401268005371, 'rewards/repeat_penalty/std': 0.0017859900835901499, 'rewards/near_duplicate_penalty/mean': 0.9990401268005371, 'rewards/near_duplicate_penalty/std': 0.0017859900835901499, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.12034481018781662, 'rewards/total_composite/std': 0.11232835054397583, 'reward': 0.12034481018781662, 'reward_std': 0.11232834309339523, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23007968068122864, 'sampling/sampling_logp_difference/max': 1.430654525756836, 'sampling/importance_sampling_ratio/min': 0.23915234208106995, 'sampling/importance_sampling_ratio/mean': 1.0516321659088135, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.193782091140747, 'clip_ratio/low_mean': 0.05039931274950504, 'clip_ratio/low_min': 0.05039931274950504, 'clip_ratio/high_mean': 0.10164953395724297, 'clip_ratio/high_max': 0.10164953395724297, 'clip_ratio/region_mean': 0.152048846706748, 'reward_total_mean': 0.12034481018781662, 'reward_meter_mean': 0.4226337671279907, 'reward_meter_std': 0.3670276403427124, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.8753880858421326, 'reward_lexical_plausibility_std': 0.2329125553369522, 'reward_repeat_penalty_mean': 0.9990401268005371, 'reward_repeat_penalty_std': 0.0017859900835901499, 'reward_near_duplicate_penalty_mean': 0.9990401268005371, 'reward_near_duplicate_penalty_std': 0.0017859900835901499, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.2749999761581421, 'reward_judge_quality_std': 0.16690459847450256, 'reward_total_composite_mean': 0.12034481018781662, 'reward_total_composite_std': 0.11232835054397583, 'epoch': 0.01}

  7%|▋         | 242/3300 [52:01<10:58:46, 12.93s/it]INFO 04-12 11:38:13 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 243/3300 [52:12<10:21:07, 12.19s/it]2026-04-12 11:38:24,149 | INFO | train_grpo_train | metrics_logged mode=train step=243

                                                     
{'loss': 0.1906, 'grad_norm': 8.332919120788574, 'learning_rate': 9.266666666666667e-06, 'num_tokens': 499880.0, 'completions/mean_length': 94.125, 'completions/min_length': 71.0, 'completions/max_length': 171.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 94.125, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 171.0, 'rewards/meter/mean': 0.8360825777053833, 'rewards/meter/std': 0.2390476018190384, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.9280816316604614, 'rewards/lexical_plausibility/std': 0.1122690737247467, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.9910176396369934, 'rewards/repeat_penalty/std': 0.011688623577356339, 'rewards/near_duplicate_penalty/mean': 0.9951965808868408, 'rewards/near_duplicate_penalty/std': 0.006244110409170389, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9958193898200989, 'rewards/distinct_2/std': 0.0118245305493474, 'rewards/total_composite/mean': 0.14266464114189148, 'rewards/total_composite/std': 0.16596995294094086, 'reward': 0.14266464114189148, 'reward_std': 0.16596995294094086, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23228609561920166, 'sampling/sampling_logp_difference/max': 1.2347612380981445, 'sampling/importance_sampling_ratio/min': 0.2909041941165924, 'sampling/importance_sampling_ratio/mean': 1.0778019428253174, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 4.365996211767197, 'clip_ratio/low_mean': 0.11267116665840149, 'clip_ratio/low_min': 0.11267116665840149, 'clip_ratio/high_mean': 0.08230546861886978, 'clip_ratio/high_max': 0.08230546861886978, 'clip_ratio/region_mean': 0.19497663527727127, 'reward_total_mean': 0.14266464114189148, 'reward_meter_mean': 0.8360825777053833, 'reward_meter_std': 0.2390476018190384, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.9280816316604614, 'reward_lexical_plausibility_std': 0.1122690737247467, 'reward_repeat_penalty_mean': 0.9910176396369934, 'reward_repeat_penalty_std': 0.011688623577356339, 'reward_near_duplicate_penalty_mean': 0.9951965808868408, 'reward_near_duplicate_penalty_std': 0.006244110409170389, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9958193898200989, 'reward_distinct_2_std': 0.0118245305493474, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.14266464114189148, 'reward_total_composite_std': 0.16596995294094086, 'epoch': 0.01}

  7%|▋         | 243/3300 [52:12<10:21:07, 12.19s/it]INFO 04-12 11:38:24 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 244/3300 [52:26<10:55:31, 12.87s/it]2026-04-12 11:38:38,605 | INFO | train_grpo_train | metrics_logged mode=train step=244

                                                     
{'loss': -0.1369, 'grad_norm': 2.5369458198547363, 'learning_rate': 9.263636363636364e-06, 'num_tokens': 502091.0, 'completions/mean_length': 385.375, 'completions/min_length': 157.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.625, 'completions/mean_terminated_length': 174.33334350585938, 'completions/min_terminated_length': 157.0, 'completions/max_terminated_length': 190.0, 'rewards/meter/mean': 0.3597410321235657, 'rewards/meter/std': 0.2850934565067291, 'rewards/count_adherence/mean': 0.6666666865348816, 'rewards/count_adherence/std': 0.3894554078578949, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/lexical_plausibility/mean': 0.6828522682189941, 'rewards/lexical_plausibility/std': 0.32094618678092957, 'rewards/judge_quality/mean': 0.11249999701976776, 'rewards/judge_quality/std': 0.1060660183429718, 'rewards/repeat_penalty/mean': 0.9637818336486816, 'rewards/repeat_penalty/std': 0.08662346750497818, 'rewards/near_duplicate_penalty/mean': 0.997008204460144, 'rewards/near_duplicate_penalty/std': 0.00376549712382257, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.998017430305481, 'rewards/distinct_2/std': 0.0056075118482112885, 'rewards/total_composite/mean': 0.048523180186748505, 'rewards/total_composite/std': 0.10372652858495712, 'reward': 0.048523180186748505, 'reward_std': 0.10372653603553772, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23052774369716644, 'sampling/sampling_logp_difference/max': 1.1155242919921875, 'sampling/importance_sampling_ratio/min': 0.3924805223941803, 'sampling/importance_sampling_ratio/mean': 1.0815635919570923, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.857869267463684, 'clip_ratio/low_mean': 0.02229299396276474, 'clip_ratio/low_min': 0.02229299396276474, 'clip_ratio/high_mean': 0.04635914973914623, 'clip_ratio/high_max': 0.04635914973914623, 'clip_ratio/region_mean': 0.06865214370191097, 'reward_total_mean': 0.048523180186748505, 'reward_meter_mean': 0.3597410321235657, 'reward_meter_std': 0.2850934565067291, 'reward_count_adherence_mean': 0.6666666865348816, 'reward_count_adherence_std': 0.3894554078578949, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_lexical_plausibility_mean': 0.6828522682189941, 'reward_lexical_plausibility_std': 0.32094618678092957, 'reward_repeat_penalty_mean': 0.9637818336486816, 'reward_repeat_penalty_std': 0.08662346750497818, 'reward_near_duplicate_penalty_mean': 0.997008204460144, 'reward_near_duplicate_penalty_std': 0.00376549712382257, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.998017430305481, 'reward_distinct_2_std': 0.0056075118482112885, 'reward_judge_quality_mean': 0.11249999701976776, 'reward_judge_quality_std': 0.1060660183429718, 'reward_total_composite_mean': 0.048523180186748505, 'reward_total_composite_std': 0.10372652858495712, 'epoch': 0.01}

  7%|▋         | 244/3300 [52:26<10:55:31, 12.87s/it]INFO 04-12 11:38:38 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 245/3300 [52:41<11:32:43, 13.61s/it]2026-04-12 11:38:53,933 | INFO | train_grpo_train | metrics_logged mode=train step=245

                                                     
{'loss': -0.1572, 'grad_norm': 2.6573731899261475, 'learning_rate': 9.260606060606062e-06, 'num_tokens': 504886.0, 'completions/mean_length': 368.375, 'completions/min_length': 184.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 224.75, 'completions/min_terminated_length': 184.0, 'completions/max_terminated_length': 306.0, 'rewards/meter/mean': 0.4802515506744385, 'rewards/meter/std': 0.18266919255256653, 'rewards/count_adherence/mean': 0.7625000476837158, 'rewards/count_adherence/std': 0.29246488213539124, 'rewards/arabic_clean/mean': 0.375, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.741745114326477, 'rewards/lexical_plausibility/std': 0.2308400571346283, 'rewards/judge_quality/mean': 0.10000000149011612, 'rewards/judge_quality/std': 0.053452253341674805, 'rewards/repeat_penalty/mean': 0.9661213755607605, 'rewards/repeat_penalty/std': 0.08737920969724655, 'rewards/near_duplicate_penalty/mean': 0.9980625510215759, 'rewards/near_duplicate_penalty/std': 0.0016265929443761706, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9993057250976562, 'rewards/distinct_2/std': 0.00196364289149642, 'rewards/total_composite/mean': 0.028423260897397995, 'rewards/total_composite/std': 0.042994480580091476, 'reward': 0.028423260897397995, 'reward_std': 0.042994480580091476, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2215947061777115, 'sampling/sampling_logp_difference/max': 1.366511344909668, 'sampling/importance_sampling_ratio/min': 0.2549950182437897, 'sampling/importance_sampling_ratio/mean': 1.059888243675232, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.8231218457221985, 'clip_ratio/low_mean': 0.015931373462080956, 'clip_ratio/low_min': 0.015931373462080956, 'clip_ratio/high_mean': 0.07494353130459785, 'clip_ratio/high_max': 0.07494353130459785, 'clip_ratio/region_mean': 0.09087490476667881, 'reward_total_mean': 0.028423260897397995, 'reward_meter_mean': 0.4802515506744385, 'reward_meter_std': 0.18266919255256653, 'reward_count_adherence_mean': 0.7625000476837158, 'reward_count_adherence_std': 0.29246488213539124, 'reward_arabic_clean_mean': 0.375, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.741745114326477, 'reward_lexical_plausibility_std': 0.2308400571346283, 'reward_repeat_penalty_mean': 0.9661213755607605, 'reward_repeat_penalty_std': 0.08737920969724655, 'reward_near_duplicate_penalty_mean': 0.9980625510215759, 'reward_near_duplicate_penalty_std': 0.0016265929443761706, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9993057250976562, 'reward_distinct_2_std': 0.00196364289149642, 'reward_judge_quality_mean': 0.10000000149011612, 'reward_judge_quality_std': 0.053452253341674805, 'reward_total_composite_mean': 0.028423260897397995, 'reward_total_composite_std': 0.042994480580091476, 'epoch': 0.01}

  7%|▋         | 245/3300 [52:41<11:32:43, 13.61s/it]INFO 04-12 11:38:54 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 246/3300 [52:51<10:36:09, 12.50s/it]2026-04-12 11:39:03,839 | INFO | train_grpo_train | metrics_logged mode=train step=246

                                                     
{'loss': 0.0746, 'grad_norm': 13.963288307189941, 'learning_rate': 9.257575757575759e-06, 'num_tokens': 506583.0, 'completions/mean_length': 54.125, 'completions/min_length': 45.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.125, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.6155509352684021, 'rewards/meter/std': 0.3467852473258972, '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/lexical_plausibility/mean': 0.9943249225616455, 'rewards/lexical_plausibility/std': 0.010853094980120659, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.9779219627380371, 'rewards/repeat_penalty/std': 0.03843917325139046, 'rewards/near_duplicate_penalty/mean': 0.9846223592758179, 'rewards/near_duplicate_penalty/std': 0.02112414687871933, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9928774833679199, 'rewards/distinct_2/std': 0.02014549821615219, 'rewards/total_composite/mean': 0.20216278731822968, 'rewards/total_composite/std': 0.1850358545780182, 'reward': 0.20216278731822968, 'reward_std': 0.185035839676857, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2559899687767029, 'sampling/sampling_logp_difference/max': 1.5820841789245605, 'sampling/importance_sampling_ratio/min': 0.20554624497890472, 'sampling/importance_sampling_ratio/mean': 1.0648609399795532, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.505162835121155, 'clip_ratio/low_mean': 0.09598811622709036, 'clip_ratio/low_min': 0.09598811622709036, 'clip_ratio/high_mean': 0.06708754226565361, 'clip_ratio/high_max': 0.06708754226565361, 'clip_ratio/region_mean': 0.16307565849274397, 'reward_total_mean': 0.20216278731822968, 'reward_meter_mean': 0.6155509352684021, 'reward_meter_std': 0.3467852473258972, '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_lexical_plausibility_mean': 0.9943249225616455, 'reward_lexical_plausibility_std': 0.010853094980120659, 'reward_repeat_penalty_mean': 0.9779219627380371, 'reward_repeat_penalty_std': 0.03843917325139046, 'reward_near_duplicate_penalty_mean': 0.9846223592758179, 'reward_near_duplicate_penalty_std': 0.02112414687871933, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9928774833679199, 'reward_distinct_2_std': 0.02014549821615219, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.20216278731822968, 'reward_total_composite_std': 0.1850358545780182, 'epoch': 0.01}

  7%|▋         | 246/3300 [52:51<10:36:09, 12.50s/it]INFO 04-12 11:39:04 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 247/3300 [53:05<10:57:05, 12.91s/it]2026-04-12 11:39:17,722 | INFO | train_grpo_train | metrics_logged mode=train step=247

                                                     
{'loss': -0.0168, 'grad_norm': 4.635673522949219, 'learning_rate': 9.254545454545454e-06, 'num_tokens': 508157.0, 'completions/mean_length': 93.75, 'completions/min_length': 31.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 34.0, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.38106468319892883, 'rewards/meter/std': 0.3492964208126068, '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/lexical_plausibility/mean': 0.9063384532928467, 'rewards/lexical_plausibility/std': 0.1857299953699112, 'rewards/judge_quality/mean': 0.42499998211860657, 'rewards/judge_quality/std': 0.19086270034313202, 'rewards/repeat_penalty/mean': 0.907705545425415, 'rewards/repeat_penalty/std': 0.15060622990131378, 'rewards/near_duplicate_penalty/mean': 0.9728717803955078, 'rewards/near_duplicate_penalty/std': 0.04505414515733719, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9896363019943237, 'rewards/distinct_2/std': 0.026975518092513084, 'rewards/total_composite/mean': 0.14492686092853546, 'rewards/total_composite/std': 0.16703473031520844, 'reward': 0.14492686092853546, 'reward_std': 0.16703473031520844, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19871190190315247, 'sampling/sampling_logp_difference/max': 1.8816604614257812, 'sampling/importance_sampling_ratio/min': 0.1523369401693344, 'sampling/importance_sampling_ratio/mean': 0.9977893233299255, 'sampling/importance_sampling_ratio/max': 1.8797191381454468, 'entropy': 1.7135359048843384, 'clip_ratio/low_mean': 0.06776629760861397, 'clip_ratio/low_min': 0.06776629760861397, 'clip_ratio/high_mean': 0.12760251387953758, 'clip_ratio/high_max': 0.12760251387953758, 'clip_ratio/region_mean': 0.19536881148815155, 'reward_total_mean': 0.14492686092853546, 'reward_meter_mean': 0.38106468319892883, 'reward_meter_std': 0.3492964208126068, '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_lexical_plausibility_mean': 0.9063384532928467, 'reward_lexical_plausibility_std': 0.1857299953699112, 'reward_repeat_penalty_mean': 0.907705545425415, 'reward_repeat_penalty_std': 0.15060622990131378, 'reward_near_duplicate_penalty_mean': 0.9728717803955078, 'reward_near_duplicate_penalty_std': 0.04505414515733719, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9896363019943237, 'reward_distinct_2_std': 0.026975518092513084, 'reward_judge_quality_mean': 0.42499998211860657, 'reward_judge_quality_std': 0.19086270034313202, 'reward_total_composite_mean': 0.14492686092853546, 'reward_total_composite_std': 0.16703473031520844, 'epoch': 0.01}

  7%|▋         | 247/3300 [53:05<10:57:05, 12.91s/it]INFO 04-12 11:39:17 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 248/3300 [53:15<10:04:39, 11.89s/it]2026-04-12 11:39:27,215 | INFO | train_grpo_train | metrics_logged mode=train step=248

                                                     
{'loss': 0.0127, 'grad_norm': 12.676898002624512, 'learning_rate': 9.251515151515152e-06, 'num_tokens': 509902.0, 'completions/mean_length': 59.125, 'completions/min_length': 47.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.125, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.6911441087722778, 'rewards/meter/std': 0.385056734085083, '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/lexical_plausibility/mean': 0.9953168630599976, 'rewards/lexical_plausibility/std': 0.008684071712195873, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.9910987615585327, 'rewards/repeat_penalty/std': 0.009570775553584099, 'rewards/near_duplicate_penalty/mean': 0.9910987615585327, 'rewards/near_duplicate_penalty/std': 0.009570775553584099, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.20069202780723572, 'rewards/total_composite/std': 0.14950935542583466, 'reward': 0.20069202780723572, 'reward_std': 0.14950934052467346, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23086237907409668, 'sampling/sampling_logp_difference/max': 1.3593578338623047, 'sampling/importance_sampling_ratio/min': 0.25682565569877625, 'sampling/importance_sampling_ratio/mean': 1.037505865097046, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.641067147254944, 'clip_ratio/low_mean': 0.07672099955379963, 'clip_ratio/low_min': 0.07672099955379963, 'clip_ratio/high_mean': 0.10969330742955208, 'clip_ratio/high_max': 0.10969330742955208, 'clip_ratio/region_mean': 0.1864143069833517, 'reward_total_mean': 0.20069202780723572, 'reward_meter_mean': 0.6911441087722778, 'reward_meter_std': 0.385056734085083, '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_lexical_plausibility_mean': 0.9953168630599976, 'reward_lexical_plausibility_std': 0.008684071712195873, 'reward_repeat_penalty_mean': 0.9910987615585327, 'reward_repeat_penalty_std': 0.009570775553584099, 'reward_near_duplicate_penalty_mean': 0.9910987615585327, 'reward_near_duplicate_penalty_std': 0.009570775553584099, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.20069202780723572, 'reward_total_composite_std': 0.14950935542583466, 'epoch': 0.01}

  8%|▊         | 248/3300 [53:15<10:04:39, 11.89s/it]INFO 04-12 11:39:27 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 249/3300 [53:30<10:50:51, 12.80s/it]2026-04-12 11:39:42,142 | INFO | train_grpo_train | metrics_logged mode=train step=249

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 9.248484848484849e-06, 'num_tokens': 513006.0, 'completions/mean_length': 410.0, 'completions/min_length': 39.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 308.0, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 491.0, 'rewards/meter/mean': 0.5463802218437195, 'rewards/meter/std': 0.3094373643398285, 'rewards/count_adherence/mean': 0.4375, 'rewards/count_adherence/std': 0.27125275135040283, 'rewards/arabic_clean/mean': 0.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.6930003762245178, 'rewards/lexical_plausibility/std': 0.2002602368593216, 'rewards/judge_quality/mean': 0.0625, 'rewards/judge_quality/std': 0.0353553406894207, 'rewards/repeat_penalty/mean': 0.9972835183143616, 'rewards/repeat_penalty/std': 0.0019352545496076345, 'rewards/near_duplicate_penalty/mean': 0.9972835183143616, 'rewards/near_duplicate_penalty/std': 0.0019352545496076345, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.0, 'rewards/total_composite/std': 0.0, 'reward': 0.0, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.19660653173923492, 'sampling/sampling_logp_difference/max': 1.0296039581298828, 'sampling/importance_sampling_ratio/min': 0.3571484088897705, 'sampling/importance_sampling_ratio/mean': 1.045520305633545, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.4071234464645386, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.0, 'reward_meter_mean': 0.5463802218437195, 'reward_meter_std': 0.3094373643398285, 'reward_count_adherence_mean': 0.4375, 'reward_count_adherence_std': 0.27125275135040283, 'reward_arabic_clean_mean': 0.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.6930003762245178, 'reward_lexical_plausibility_std': 0.2002602368593216, 'reward_repeat_penalty_mean': 0.9972835183143616, 'reward_repeat_penalty_std': 0.0019352545496076345, 'reward_near_duplicate_penalty_mean': 0.9972835183143616, 'reward_near_duplicate_penalty_std': 0.0019352545496076345, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.0625, 'reward_judge_quality_std': 0.0353553406894207, 'reward_total_composite_mean': 0.0, 'reward_total_composite_std': 0.0, 'epoch': 0.01}

  8%|▊         | 249/3300 [53:30<10:50:51, 12.80s/it]INFO 04-12 11:39:42 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 250/3300 [53:40<10:10:14, 12.00s/it]2026-04-12 11:39:52,295 | INFO | train_grpo_train | metrics_logged mode=train step=250

                                                     
{'loss': 0.026, 'grad_norm': 14.034645080566406, 'learning_rate': 9.245454545454546e-06, 'num_tokens': 514633.0, 'completions/mean_length': 50.375, 'completions/min_length': 46.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.375, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.37017858028411865, 'rewards/meter/std': 0.2679588198661804, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9791666269302368, 'rewards/lexical_plausibility/std': 0.0589255727827549, 'rewards/judge_quality/mean': 0.4462500214576721, 'rewards/judge_quality/std': 0.2173829823732376, 'rewards/repeat_penalty/mean': 0.9731531143188477, 'rewards/repeat_penalty/std': 0.02014872431755066, 'rewards/near_duplicate_penalty/mean': 0.9810901880264282, 'rewards/near_duplicate_penalty/std': 0.008582751266658306, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9919871687889099, 'rewards/distinct_2/std': 0.02266368828713894, 'rewards/total_composite/mean': 0.15338610112667084, 'rewards/total_composite/std': 0.10794755816459656, 'reward': 0.15338610112667084, 'reward_std': 0.10794755071401596, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23086252808570862, 'sampling/sampling_logp_difference/max': 3.903805732727051, 'sampling/importance_sampling_ratio/min': 0.02016502246260643, 'sampling/importance_sampling_ratio/mean': 1.0236157178878784, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.3722375631332397, 'clip_ratio/low_mean': 0.09338569827377796, 'clip_ratio/low_min': 0.09338569827377796, 'clip_ratio/high_mean': 0.10044643003493547, 'clip_ratio/high_max': 0.10044643003493547, 'clip_ratio/region_mean': 0.19383212830871344, 'reward_total_mean': 0.15338610112667084, 'reward_meter_mean': 0.37017858028411865, 'reward_meter_std': 0.2679588198661804, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9791666269302368, 'reward_lexical_plausibility_std': 0.0589255727827549, 'reward_repeat_penalty_mean': 0.9731531143188477, 'reward_repeat_penalty_std': 0.02014872431755066, 'reward_near_duplicate_penalty_mean': 0.9810901880264282, 'reward_near_duplicate_penalty_std': 0.008582751266658306, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9919871687889099, 'reward_distinct_2_std': 0.02266368828713894, 'reward_judge_quality_mean': 0.4462500214576721, 'reward_judge_quality_std': 0.2173829823732376, 'reward_total_composite_mean': 0.15338610112667084, 'reward_total_composite_std': 0.10794755816459656, 'epoch': 0.01}

  8%|▊         | 250/3300 [53:40<10:10:14, 12.00s/it]INFO 04-12 11:39:52 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


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

 15%|█▌        | 2/13 [00:12<01:07,  6.10s/it]

 23%|██▎       | 3/13 [00:21<01:13,  7.36s/it]

 31%|███       | 4/13 [00:33<01:21,  9.06s/it]

 38%|███▊      | 5/13 [00:47<01:27, 10.99s/it]

 46%|████▌     | 6/13 [01:00<01:19, 11.41s/it]

 54%|█████▍    | 7/13 [01:13<01:12, 12.02s/it]

 62%|██████▏   | 8/13 [01:25<01:00, 12.17s/it]

 69%|██████▉   | 9/13 [01:38<00:49, 12.33s/it]

 77%|███████▋  | 10/13 [01:49<00:36, 12.02s/it]

 85%|████████▍ | 11/13 [02:03<00:24, 12.41s/it]

 92%|█████████▏| 12/13 [02:15<00:12, 12.26s/it]

100%|██████████| 13/13 [02:28<00:00, 12.69s/it]2026-04-12 11:42:35,437 | INFO | train_grpo_train | metrics_logged mode=eval step=250

                                                     

                                               
{'eval_loss': nan, 'eval_runtime': 163.0871, 'eval_samples_per_second': 0.638, 'eval_steps_per_second': 0.08, 'eval_num_tokens': 514633.0, 'eval_completions/mean_length': 190.29807692307693, 'eval_completions/min_length': 41.07692307692308, 'eval_completions/max_length': 459.84615384615387, 'eval_completions/clipped_ratio': 0.18269230769230768, 'eval_completions/mean_terminated_length': 120.03590158315805, 'eval_completions/min_terminated_length': 41.07692307692308, 'eval_completions/max_terminated_length': 247.30769230769232, 'eval_rewards/meter/mean': 0.35175654177482313, 'eval_rewards/meter/std': 0.33574836070720965, 'eval_rewards/count_adherence/mean': 0.9094114349438593, 'eval_rewards/count_adherence/std': 0.15080647428448385, 'eval_rewards/arabic_clean/mean': 0.7403846153846154, 'eval_rewards/arabic_clean/std': 0.3887776663670173, 'eval_rewards/lexical_plausibility/mean': 0.9016158397381122, 'eval_rewards/lexical_plausibility/std': 0.1482962743880657, 'eval_rewards/judge_quality/mean': 0.319519233245116, 'eval_rewards/judge_quality/std': 0.20452269395956627, 'eval_rewards/repeat_penalty/mean': 0.9738312730422387, 'eval_rewards/repeat_penalty/std': 0.04887550560614237, 'eval_rewards/near_duplicate_penalty/mean': 0.9904315425799444, 'eval_rewards/near_duplicate_penalty/std': 0.01207644809395648, 'eval_rewards/opening_diversity/mean': 0.9927884615384616, 'eval_rewards/opening_diversity/std': 0.020397310646680687, 'eval_rewards/distinct_2/mean': 0.9907622520740216, 'eval_rewards/distinct_2/std': 0.020605855603487447, 'eval_rewards/total_composite/mean': 0.11678541100655611, 'eval_rewards/total_composite/std': 0.14927118780234686, 'eval_reward': 0.11678541100655611, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.1658907303443322, 'eval_sampling/sampling_logp_difference/max': 1.1860259129450872, 'eval_sampling/importance_sampling_ratio/min': 0.3094748098116655, 'eval_sampling/importance_sampling_ratio/mean': 1.0527949975087092, 'eval_sampling/importance_sampling_ratio/max': 1.7494422655839186, 'eval_entropy': 3.38864253117488, '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': 0.11678541100655611, 'eval_reward_meter_mean': 0.35175654177482313, 'eval_reward_meter_std': 0.33574836070720965, 'eval_reward_count_adherence_mean': 0.9094114349438593, 'eval_reward_count_adherence_std': 0.15080647428448385, 'eval_reward_arabic_clean_mean': 0.7403846153846154, 'eval_reward_arabic_clean_std': 0.3887776663670173, 'eval_reward_lexical_plausibility_mean': 0.9016158397381122, 'eval_reward_lexical_plausibility_std': 0.1482962743880657, 'eval_reward_repeat_penalty_mean': 0.9738312730422387, 'eval_reward_repeat_penalty_std': 0.04887550560614237, 'eval_reward_near_duplicate_penalty_mean': 0.9904315425799444, 'eval_reward_near_duplicate_penalty_std': 0.01207644809395648, 'eval_reward_opening_diversity_mean': 0.9927884615384616, 'eval_reward_opening_diversity_std': 0.020397310646680687, 'eval_reward_distinct_2_mean': 0.9907622520740216, 'eval_reward_distinct_2_std': 0.020605855603487447, 'eval_reward_judge_quality_mean': 0.319519233245116, 'eval_reward_judge_quality_std': 0.20452269395956627, 'eval_reward_total_composite_mean': 0.11678541100655611, 'eval_reward_total_composite_std': 0.14927118780234686, 'epoch': 0.01}

  8%|▊         | 250/3300 [56:23<10:10:14, 12.00s/it]

100%|██████████| 13/13 [02:28<00:00, 12.69s/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(

  8%|▊         | 251/3300 [56:34<51:31:00, 60.83s/it]2026-04-12 11:42:47,038 | INFO | train_grpo_train | metrics_logged mode=train step=251

                                                     
{'loss': 0.1223, 'grad_norm': 15.447683334350586, 'learning_rate': 9.242424242424244e-06, 'num_tokens': 516312.0, 'completions/mean_length': 42.875, 'completions/min_length': 36.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.875, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.9646941423416138, 'rewards/meter/std': 0.036442022770643234, '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/lexical_plausibility/mean': 0.9295889139175415, 'rewards/lexical_plausibility/std': 0.12642835080623627, 'rewards/judge_quality/mean': 0.39625000953674316, 'rewards/judge_quality/std': 0.2585087716579437, 'rewards/repeat_penalty/mean': 0.9976711273193359, 'rewards/repeat_penalty/std': 0.004313847981393337, 'rewards/near_duplicate_penalty/mean': 0.9976711273193359, 'rewards/near_duplicate_penalty/std': 0.004313847981393337, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3644532263278961, 'rewards/total_composite/std': 0.27539464831352234, 'reward': 0.3644532263278961, 'reward_std': 0.27539464831352234, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22599036991596222, 'sampling/sampling_logp_difference/max': 1.4986276626586914, 'sampling/importance_sampling_ratio/min': 0.22343659400939941, 'sampling/importance_sampling_ratio/mean': 1.043885350227356, 'sampling/importance_sampling_ratio/max': 1.9157660007476807, 'entropy': 3.080083429813385, 'clip_ratio/low_mean': 0.07370169088244438, 'clip_ratio/low_min': 0.07370169088244438, 'clip_ratio/high_mean': 0.14847705326974392, 'clip_ratio/high_max': 0.14847705326974392, 'clip_ratio/region_mean': 0.2221787441521883, 'reward_total_mean': 0.3644532263278961, 'reward_meter_mean': 0.9646941423416138, 'reward_meter_std': 0.036442022770643234, '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_lexical_plausibility_mean': 0.9295889139175415, 'reward_lexical_plausibility_std': 0.12642835080623627, 'reward_repeat_penalty_mean': 0.9976711273193359, 'reward_repeat_penalty_std': 0.004313847981393337, 'reward_near_duplicate_penalty_mean': 0.9976711273193359, 'reward_near_duplicate_penalty_std': 0.004313847981393337, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.39625000953674316, 'reward_judge_quality_std': 0.2585087716579437, 'reward_total_composite_mean': 0.3644532263278961, 'reward_total_composite_std': 0.27539464831352234, 'epoch': 0.01}

  8%|▊         | 251/3300 [56:35<51:31:00, 60.83s/it]INFO 04-12 11:42:47 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 252/3300 [56:44<38:27:01, 45.41s/it]2026-04-12 11:42:56,488 | INFO | train_grpo_train | metrics_logged mode=train step=252

                                                     
{'loss': 0.1343, 'grad_norm': 13.339502334594727, 'learning_rate': 9.23939393939394e-06, 'num_tokens': 517916.0, 'completions/mean_length': 47.5, 'completions/min_length': 40.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.5, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.7088395953178406, 'rewards/meter/std': 0.34888893365859985, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9529064893722534, 'rewards/lexical_plausibility/std': 0.09228985756635666, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.23260942101478577, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.24300158023834229, 'rewards/total_composite/std': 0.26558998227119446, 'reward': 0.24300158023834229, 'reward_std': 0.26558998227119446, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22465598583221436, 'sampling/sampling_logp_difference/max': 1.2926592826843262, 'sampling/importance_sampling_ratio/min': 0.2745397090911865, 'sampling/importance_sampling_ratio/mean': 1.0339922904968262, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.0143211483955383, 'clip_ratio/low_mean': 0.10859714634716511, 'clip_ratio/low_min': 0.10859714634716511, 'clip_ratio/high_mean': 0.08645833656191826, 'clip_ratio/high_max': 0.08645833656191826, 'clip_ratio/region_mean': 0.19505548290908337, 'reward_total_mean': 0.24300158023834229, 'reward_meter_mean': 0.7088395953178406, 'reward_meter_std': 0.34888893365859985, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9529064893722534, 'reward_lexical_plausibility_std': 0.09228985756635666, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.23260942101478577, 'reward_total_composite_mean': 0.24300158023834229, 'reward_total_composite_std': 0.26558998227119446, 'epoch': 0.01}

  8%|▊         | 252/3300 [56:44<38:27:01, 45.41s/it]INFO 04-12 11:42:56 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 253/3300 [56:59<30:48:40, 36.40s/it]2026-04-12 11:43:11,866 | INFO | train_grpo_train | metrics_logged mode=train step=253

                                                     
{'loss': -0.073, 'grad_norm': 2.8028078079223633, 'learning_rate': 9.236363636363636e-06, 'num_tokens': 519973.0, 'completions/mean_length': 271.125, 'completions/min_length': 98.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 126.5999984741211, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 215.0, 'rewards/meter/mean': 0.44935062527656555, 'rewards/meter/std': 0.3874340057373047, 'rewards/count_adherence/mean': 0.8958333134651184, 'rewards/count_adherence/std': 0.12400396913290024, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/lexical_plausibility/mean': 0.8119495511054993, 'rewards/lexical_plausibility/std': 0.2038293480873108, 'rewards/judge_quality/mean': 0.23749999701976776, 'rewards/judge_quality/std': 0.20310096442699432, 'rewards/repeat_penalty/mean': 0.9870443344116211, 'rewards/repeat_penalty/std': 0.018468057736754417, 'rewards/near_duplicate_penalty/mean': 0.9940108060836792, 'rewards/near_duplicate_penalty/std': 0.007382832467556, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9929375052452087, 'rewards/distinct_2/std': 0.013083843514323235, 'rewards/total_composite/mean': 0.1686832755804062, 'rewards/total_composite/std': 0.19913244247436523, 'reward': 0.1686832755804062, 'reward_std': 0.19913245737552643, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20941412448883057, 'sampling/sampling_logp_difference/max': 1.2020454406738281, 'sampling/importance_sampling_ratio/min': 0.300578773021698, 'sampling/importance_sampling_ratio/mean': 1.0455708503723145, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6216647922992706, 'clip_ratio/low_mean': 0.030135659500956535, 'clip_ratio/low_min': 0.030135659500956535, 'clip_ratio/high_mean': 0.08487984724342823, 'clip_ratio/high_max': 0.08487984724342823, 'clip_ratio/region_mean': 0.11501550674438477, 'reward_total_mean': 0.1686832755804062, 'reward_meter_mean': 0.44935062527656555, 'reward_meter_std': 0.3874340057373047, 'reward_count_adherence_mean': 0.8958333134651184, 'reward_count_adherence_std': 0.12400396913290024, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_lexical_plausibility_mean': 0.8119495511054993, 'reward_lexical_plausibility_std': 0.2038293480873108, 'reward_repeat_penalty_mean': 0.9870443344116211, 'reward_repeat_penalty_std': 0.018468057736754417, 'reward_near_duplicate_penalty_mean': 0.9940108060836792, 'reward_near_duplicate_penalty_std': 0.007382832467556, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9929375052452087, 'reward_distinct_2_std': 0.013083843514323235, 'reward_judge_quality_mean': 0.23749999701976776, 'reward_judge_quality_std': 0.20310096442699432, 'reward_total_composite_mean': 0.1686832755804062, 'reward_total_composite_std': 0.19913244247436523, 'epoch': 0.01}

  8%|▊         | 253/3300 [56:59<30:48:40, 36.40s/it]INFO 04-12 11:43:12 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 254/3300 [57:09<23:54:21, 28.25s/it]2026-04-12 11:43:21,106 | INFO | train_grpo_train | metrics_logged mode=train step=254

                                                     
{'loss': 0.1091, 'grad_norm': 22.059316635131836, 'learning_rate': 9.233333333333334e-06, 'num_tokens': 521365.0, 'completions/mean_length': 25.0, 'completions/min_length': 21.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.0, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.4769478440284729, 'rewards/meter/std': 0.37168920040130615, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9941449165344238, 'rewards/lexical_plausibility/std': 0.016560783609747887, 'rewards/judge_quality/mean': 0.5337499976158142, 'rewards/judge_quality/std': 0.17104199528694153, 'rewards/repeat_penalty/mean': 0.9345779418945312, 'rewards/repeat_penalty/std': 0.14903269708156586, 'rewards/near_duplicate_penalty/mean': 0.9735829830169678, 'rewards/near_duplicate_penalty/std': 0.04036236181855202, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9825174808502197, 'rewards/distinct_2/std': 0.049448031932115555, 'rewards/total_composite/mean': 0.25110459327697754, 'rewards/total_composite/std': 0.2017320841550827, 'reward': 0.25110459327697754, 'reward_std': 0.2017320990562439, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17039695382118225, 'sampling/sampling_logp_difference/max': 1.5042152404785156, 'sampling/importance_sampling_ratio/min': 0.22219161689281464, 'sampling/importance_sampling_ratio/mean': 1.0191543102264404, 'sampling/importance_sampling_ratio/max': 1.8768799304962158, 'entropy': 1.199865683913231, 'clip_ratio/low_mean': 0.09119006153196096, 'clip_ratio/low_min': 0.09119006153196096, 'clip_ratio/high_mean': 0.08217719849199057, 'clip_ratio/high_max': 0.08217719849199057, 'clip_ratio/region_mean': 0.17336726002395153, 'reward_total_mean': 0.25110459327697754, 'reward_meter_mean': 0.4769478440284729, 'reward_meter_std': 0.37168920040130615, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9941449165344238, 'reward_lexical_plausibility_std': 0.016560783609747887, 'reward_repeat_penalty_mean': 0.9345779418945312, 'reward_repeat_penalty_std': 0.14903269708156586, 'reward_near_duplicate_penalty_mean': 0.9735829830169678, 'reward_near_duplicate_penalty_std': 0.04036236181855202, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9825174808502197, 'reward_distinct_2_std': 0.049448031932115555, 'reward_judge_quality_mean': 0.5337499976158142, 'reward_judge_quality_std': 0.17104199528694153, 'reward_total_composite_mean': 0.25110459327697754, 'reward_total_composite_std': 0.2017320841550827, 'epoch': 0.01}

  8%|▊         | 254/3300 [57:09<23:54:21, 28.25s/it]INFO 04-12 11:43:21 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 255/3300 [57:24<20:45:17, 24.54s/it]2026-04-12 11:43:36,972 | INFO | train_grpo_train | metrics_logged mode=train step=255

                                                     
{'loss': -0.1652, 'grad_norm': 3.5489718914031982, 'learning_rate': 9.23030303030303e-06, 'num_tokens': 524589.0, 'completions/mean_length': 308.0, 'completions/min_length': 213.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 240.0, 'completions/min_terminated_length': 213.0, 'completions/max_terminated_length': 284.0, 'rewards/meter/mean': 0.6008556485176086, 'rewards/meter/std': 0.15912260115146637, 'rewards/count_adherence/mean': 0.8541666269302368, 'rewards/count_adherence/std': 0.16517187654972076, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.8297012448310852, 'rewards/lexical_plausibility/std': 0.1053345650434494, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.09258200973272324, 'rewards/repeat_penalty/mean': 0.9708657264709473, 'rewards/repeat_penalty/std': 0.024701109156012535, 'rewards/near_duplicate_penalty/mean': 0.9894903898239136, 'rewards/near_duplicate_penalty/std': 0.007003883831202984, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9811153411865234, 'rewards/distinct_2/std': 0.020878588780760765, 'rewards/total_composite/mean': 0.0616358257830143, 'rewards/total_composite/std': 0.06312758475542068, 'reward': 0.0616358257830143, 'reward_std': 0.06312758475542068, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21726159751415253, 'sampling/sampling_logp_difference/max': 1.3737688064575195, 'sampling/importance_sampling_ratio/min': 0.2531510591506958, 'sampling/importance_sampling_ratio/mean': 1.0538618564605713, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.8067986369132996, 'clip_ratio/low_mean': 0.027880338951945305, 'clip_ratio/low_min': 0.027880338951945305, 'clip_ratio/high_mean': 0.08326276205480099, 'clip_ratio/high_max': 0.08326276205480099, 'clip_ratio/region_mean': 0.11114310100674629, 'reward_total_mean': 0.0616358257830143, 'reward_meter_mean': 0.6008556485176086, 'reward_meter_std': 0.15912260115146637, 'reward_count_adherence_mean': 0.8541666269302368, 'reward_count_adherence_std': 0.16517187654972076, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.8297012448310852, 'reward_lexical_plausibility_std': 0.1053345650434494, 'reward_repeat_penalty_mean': 0.9708657264709473, 'reward_repeat_penalty_std': 0.024701109156012535, 'reward_near_duplicate_penalty_mean': 0.9894903898239136, 'reward_near_duplicate_penalty_std': 0.007003883831202984, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9811153411865234, 'reward_distinct_2_std': 0.020878588780760765, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.09258200973272324, 'reward_total_composite_mean': 0.0616358257830143, 'reward_total_composite_std': 0.06312758475542068, 'epoch': 0.01}

  8%|▊         | 255/3300 [57:24<20:45:17, 24.54s/it]INFO 04-12 11:43:37 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 256/3300 [57:34<16:52:23, 19.96s/it]2026-04-12 11:43:46,236 | INFO | train_grpo_train | metrics_logged mode=train step=256

                                                     
{'loss': 0.0754, 'grad_norm': 27.22827911376953, 'learning_rate': 9.227272727272728e-06, 'num_tokens': 525915.0, 'completions/mean_length': 17.75, 'completions/min_length': 15.0, 'completions/max_length': 24.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 17.75, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.7467088103294373, 'rewards/meter/std': 0.3884599804878235, '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/lexical_plausibility/mean': 0.8452818393707275, 'rewards/lexical_plausibility/std': 0.18745341897010803, 'rewards/judge_quality/mean': 0.3800000250339508, 'rewards/judge_quality/std': 0.3489985764026642, 'rewards/repeat_penalty/mean': 0.729301929473877, 'rewards/repeat_penalty/std': 0.058542944490909576, 'rewards/near_duplicate_penalty/mean': 0.9724025726318359, 'rewards/near_duplicate_penalty/std': 0.07805725187063217, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.22989892959594727, 'rewards/total_composite/std': 0.2969686985015869, 'reward': 0.22989892959594727, 'reward_std': 0.2969686686992645, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23513489961624146, 'sampling/sampling_logp_difference/max': 1.5084762573242188, 'sampling/importance_sampling_ratio/min': 0.2212468534708023, 'sampling/importance_sampling_ratio/mean': 1.0271424055099487, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.3353227227926254, 'clip_ratio/low_mean': 0.15449346834793687, 'clip_ratio/low_min': 0.15449346834793687, 'clip_ratio/high_mean': 0.05964912474155426, 'clip_ratio/high_max': 0.05964912474155426, 'clip_ratio/region_mean': 0.21414259308949113, 'reward_total_mean': 0.22989892959594727, 'reward_meter_mean': 0.7467088103294373, 'reward_meter_std': 0.3884599804878235, '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_lexical_plausibility_mean': 0.8452818393707275, 'reward_lexical_plausibility_std': 0.18745341897010803, 'reward_repeat_penalty_mean': 0.729301929473877, 'reward_repeat_penalty_std': 0.058542944490909576, 'reward_near_duplicate_penalty_mean': 0.9724025726318359, 'reward_near_duplicate_penalty_std': 0.07805725187063217, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3800000250339508, 'reward_judge_quality_std': 0.3489985764026642, 'reward_total_composite_mean': 0.22989892959594727, 'reward_total_composite_std': 0.2969686985015869, 'epoch': 0.01}

  8%|▊         | 256/3300 [57:34<16:52:23, 19.96s/it]INFO 04-12 11:43:46 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 257/3300 [57:49<15:47:23, 18.68s/it]2026-04-12 11:44:01,940 | INFO | train_grpo_train | metrics_logged mode=train step=257

                                                     
{'loss': -0.0162, 'grad_norm': 5.925915718078613, 'learning_rate': 9.224242424242424e-06, 'num_tokens': 527771.0, 'completions/mean_length': 118.0, 'completions/min_length': 48.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 61.71428680419922, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.5321698188781738, 'rewards/meter/std': 0.3322257101535797, '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/lexical_plausibility/mean': 0.9015945196151733, 'rewards/lexical_plausibility/std': 0.2712819278240204, 'rewards/judge_quality/mean': 0.42124998569488525, 'rewards/judge_quality/std': 0.24032345414161682, 'rewards/repeat_penalty/mean': 0.9819114208221436, 'rewards/repeat_penalty/std': 0.017748935148119926, 'rewards/near_duplicate_penalty/mean': 0.9819114208221436, 'rewards/near_duplicate_penalty/std': 0.017748935148119926, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.22129103541374207, 'rewards/total_composite/std': 0.2818167209625244, 'reward': 0.22129103541374207, 'reward_std': 0.2818167209625244, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22934019565582275, 'sampling/sampling_logp_difference/max': 1.6250391006469727, 'sampling/importance_sampling_ratio/min': 0.19690397381782532, 'sampling/importance_sampling_ratio/mean': 1.0515106916427612, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2498981803655624, 'clip_ratio/low_mean': 0.09621335752308369, 'clip_ratio/low_min': 0.09621335752308369, 'clip_ratio/high_mean': 0.07149841822683811, 'clip_ratio/high_max': 0.07149841822683811, 'clip_ratio/region_mean': 0.1677117757499218, 'reward_total_mean': 0.22129103541374207, 'reward_meter_mean': 0.5321698188781738, 'reward_meter_std': 0.3322257101535797, '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_lexical_plausibility_mean': 0.9015945196151733, 'reward_lexical_plausibility_std': 0.2712819278240204, 'reward_repeat_penalty_mean': 0.9819114208221436, 'reward_repeat_penalty_std': 0.017748935148119926, 'reward_near_duplicate_penalty_mean': 0.9819114208221436, 'reward_near_duplicate_penalty_std': 0.017748935148119926, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.42124998569488525, 'reward_judge_quality_std': 0.24032345414161682, 'reward_total_composite_mean': 0.22129103541374207, 'reward_total_composite_std': 0.2818167209625244, 'epoch': 0.01}

  8%|▊         | 257/3300 [57:49<15:47:23, 18.68s/it]INFO 04-12 11:44:02 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 258/3300 [58:04<14:48:24, 17.52s/it]2026-04-12 11:44:16,761 | INFO | train_grpo_train | metrics_logged mode=train step=258

                                                     
{'loss': -0.0621, 'grad_norm': 5.228208541870117, 'learning_rate': 9.221212121212123e-06, 'num_tokens': 529270.0, 'completions/mean_length': 95.375, 'completions/min_length': 31.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 35.85714340209961, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.530077338218689, 'rewards/meter/std': 0.4401782155036926, '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/lexical_plausibility/mean': 0.9359341859817505, 'rewards/lexical_plausibility/std': 0.18120546638965607, 'rewards/judge_quality/mean': 0.5087500214576721, 'rewards/judge_quality/std': 0.2959941029548645, 'rewards/repeat_penalty/mean': 0.9931499361991882, 'rewards/repeat_penalty/std': 0.016404850408434868, 'rewards/near_duplicate_penalty/mean': 0.9931499361991882, 'rewards/near_duplicate_penalty/std': 0.016404850408434868, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3076081871986389, 'rewards/total_composite/std': 0.2989545464515686, 'reward': 0.3076081871986389, 'reward_std': 0.2989545464515686, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21388907730579376, 'sampling/sampling_logp_difference/max': 1.5502357482910156, 'sampling/importance_sampling_ratio/min': 0.212197944521904, 'sampling/importance_sampling_ratio/mean': 1.0432251691818237, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9004587233066559, 'clip_ratio/low_mean': 0.04779411852359772, 'clip_ratio/low_min': 0.04779411852359772, 'clip_ratio/high_mean': 0.1295657865703106, 'clip_ratio/high_max': 0.1295657865703106, 'clip_ratio/region_mean': 0.1773599050939083, 'reward_total_mean': 0.3076081871986389, 'reward_meter_mean': 0.530077338218689, 'reward_meter_std': 0.4401782155036926, '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_lexical_plausibility_mean': 0.9359341859817505, 'reward_lexical_plausibility_std': 0.18120546638965607, 'reward_repeat_penalty_mean': 0.9931499361991882, 'reward_repeat_penalty_std': 0.016404850408434868, 'reward_near_duplicate_penalty_mean': 0.9931499361991882, 'reward_near_duplicate_penalty_std': 0.016404850408434868, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5087500214576721, 'reward_judge_quality_std': 0.2959941029548645, 'reward_total_composite_mean': 0.3076081871986389, 'reward_total_composite_std': 0.2989545464515686, 'epoch': 0.01}

  8%|▊         | 258/3300 [58:04<14:48:24, 17.52s/it]INFO 04-12 11:44:17 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 259/3300 [58:14<12:56:19, 15.32s/it]2026-04-12 11:44:26,931 | INFO | train_grpo_train | metrics_logged mode=train step=259

                                                     
{'loss': 0.0289, 'grad_norm': 12.575897216796875, 'learning_rate': 9.21818181818182e-06, 'num_tokens': 531301.0, 'completions/mean_length': 88.875, 'completions/min_length': 75.0, 'completions/max_length': 100.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 88.875, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 100.0, 'rewards/meter/mean': 0.5499953031539917, 'rewards/meter/std': 0.39406150579452515, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.42124998569488525, 'rewards/judge_quality/std': 0.22177450358867645, 'rewards/repeat_penalty/mean': 0.983332633972168, 'rewards/repeat_penalty/std': 0.021853748708963394, 'rewards/near_duplicate_penalty/mean': 0.9898728728294373, 'rewards/near_duplicate_penalty/std': 0.009024342522025108, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9933686852455139, 'rewards/distinct_2/std': 0.01875614933669567, 'rewards/total_composite/mean': 0.26504188776016235, 'rewards/total_composite/std': 0.25803911685943604, 'reward': 0.26504188776016235, 'reward_std': 0.25803911685943604, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16319558024406433, 'sampling/sampling_logp_difference/max': 2.4158849716186523, 'sampling/importance_sampling_ratio/min': 0.08928828686475754, 'sampling/importance_sampling_ratio/mean': 1.011214017868042, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9410185590386391, 'clip_ratio/low_mean': 0.06891975738108158, 'clip_ratio/low_min': 0.06891975738108158, 'clip_ratio/high_mean': 0.053794910199940205, 'clip_ratio/high_max': 0.053794910199940205, 'clip_ratio/region_mean': 0.12271466758102179, 'reward_total_mean': 0.26504188776016235, 'reward_meter_mean': 0.5499953031539917, 'reward_meter_std': 0.39406150579452515, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.983332633972168, 'reward_repeat_penalty_std': 0.021853748708963394, 'reward_near_duplicate_penalty_mean': 0.9898728728294373, 'reward_near_duplicate_penalty_std': 0.009024342522025108, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9933686852455139, 'reward_distinct_2_std': 0.01875614933669567, 'reward_judge_quality_mean': 0.42124998569488525, 'reward_judge_quality_std': 0.22177450358867645, 'reward_total_composite_mean': 0.26504188776016235, 'reward_total_composite_std': 0.25803911685943604, 'epoch': 0.01}

  8%|▊         | 259/3300 [58:14<12:56:19, 15.32s/it]INFO 04-12 11:44:27 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 260/3300 [58:24<11:31:08, 13.64s/it]2026-04-12 11:44:36,662 | INFO | train_grpo_train | metrics_logged mode=train step=260

                                                     
{'loss': -0.0745, 'grad_norm': 14.188420295715332, 'learning_rate': 9.215151515151515e-06, 'num_tokens': 532931.0, 'completions/mean_length': 37.75, 'completions/min_length': 22.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.75, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.6892690658569336, 'rewards/meter/std': 0.3877565860748291, '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/lexical_plausibility/mean': 0.97265625, 'rewards/lexical_plausibility/std': 0.07733980566263199, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.11649646610021591, 'rewards/repeat_penalty/mean': 0.9641187191009521, 'rewards/repeat_penalty/std': 0.08692087978124619, 'rewards/near_duplicate_penalty/mean': 0.9953687191009521, 'rewards/near_duplicate_penalty/std': 0.008575868792831898, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.21461795270442963, 'rewards/total_composite/std': 0.16348567605018616, 'reward': 0.21461795270442963, 'reward_std': 0.16348569095134735, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22446295619010925, 'sampling/sampling_logp_difference/max': 1.6275043487548828, 'sampling/importance_sampling_ratio/min': 0.1964191496372223, 'sampling/importance_sampling_ratio/mean': 1.0513017177581787, 'sampling/importance_sampling_ratio/max': 1.9789457321166992, 'entropy': 2.5972491651773453, 'clip_ratio/low_mean': 0.06940420810133219, 'clip_ratio/low_min': 0.06940420810133219, 'clip_ratio/high_mean': 0.08271594252437353, 'clip_ratio/high_max': 0.08271594252437353, 'clip_ratio/region_mean': 0.15212015062570572, 'reward_total_mean': 0.21461795270442963, 'reward_meter_mean': 0.6892690658569336, 'reward_meter_std': 0.3877565860748291, '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_lexical_plausibility_mean': 0.97265625, 'reward_lexical_plausibility_std': 0.07733980566263199, 'reward_repeat_penalty_mean': 0.9641187191009521, 'reward_repeat_penalty_std': 0.08692087978124619, 'reward_near_duplicate_penalty_mean': 0.9953687191009521, 'reward_near_duplicate_penalty_std': 0.008575868792831898, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.11649646610021591, 'reward_total_composite_mean': 0.21461795270442963, 'reward_total_composite_std': 0.16348567605018616, 'epoch': 0.01}

  8%|▊         | 260/3300 [58:24<11:31:08, 13.64s/it]INFO 04-12 11:44:36 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 261/3300 [58:33<10:25:30, 12.35s/it]2026-04-12 11:44:45,998 | INFO | train_grpo_train | metrics_logged mode=train step=261

                                                     
{'loss': 0.0094, 'grad_norm': 29.37902069091797, 'learning_rate': 9.212121212121213e-06, 'num_tokens': 534267.0, 'completions/mean_length': 21.0, 'completions/min_length': 16.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.0, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.7790894508361816, 'rewards/meter/std': 0.3281017243862152, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9209558963775635, 'rewards/lexical_plausibility/std': 0.22357052564620972, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.08864051848649979, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.23637709021568298, 'rewards/total_composite/std': 0.11187026649713516, 'reward': 0.23637709021568298, 'reward_std': 0.11187026649713516, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21879105269908905, 'sampling/sampling_logp_difference/max': 1.655721664428711, 'sampling/importance_sampling_ratio/min': 0.19095420837402344, 'sampling/importance_sampling_ratio/mean': 1.0388354063034058, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8359774500131607, 'clip_ratio/low_mean': 0.04009103728458285, 'clip_ratio/low_min': 0.04009103728458285, 'clip_ratio/high_mean': 0.1419555712491274, 'clip_ratio/high_max': 0.1419555712491274, 'clip_ratio/region_mean': 0.18204660853371024, 'reward_total_mean': 0.23637709021568298, 'reward_meter_mean': 0.7790894508361816, 'reward_meter_std': 0.3281017243862152, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9209558963775635, 'reward_lexical_plausibility_std': 0.22357052564620972, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.08864051848649979, 'reward_total_composite_mean': 0.23637709021568298, 'reward_total_composite_std': 0.11187026649713516, 'epoch': 0.01}

  8%|▊         | 261/3300 [58:33<10:25:30, 12.35s/it]INFO 04-12 11:44:46 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 262/3300 [58:44<10:04:31, 11.94s/it]2026-04-12 11:44:56,980 | INFO | train_grpo_train | metrics_logged mode=train step=262

                                                     
{'loss': 0.0591, 'grad_norm': 12.937562942504883, 'learning_rate': 9.20909090909091e-06, 'num_tokens': 536511.0, 'completions/mean_length': 85.5, 'completions/min_length': 74.0, 'completions/max_length': 93.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 85.5, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.5915085673332214, 'rewards/meter/std': 0.37578529119491577, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9946610331535339, 'rewards/lexical_plausibility/std': 0.012002813629806042, 'rewards/judge_quality/mean': 0.42499998211860657, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.985454797744751, 'rewards/repeat_penalty/std': 0.01338482741266489, 'rewards/near_duplicate_penalty/mean': 0.9898806810379028, 'rewards/near_duplicate_penalty/std': 0.004475778434425592, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9955277442932129, 'rewards/distinct_2/std': 0.012649492360651493, 'rewards/total_composite/mean': 0.2553885281085968, 'rewards/total_composite/std': 0.16473935544490814, 'reward': 0.2553885281085968, 'reward_std': 0.16473934054374695, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18291543424129486, 'sampling/sampling_logp_difference/max': 1.422365665435791, 'sampling/importance_sampling_ratio/min': 0.2411428987979889, 'sampling/importance_sampling_ratio/mean': 1.0286941528320312, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8144361525774002, 'clip_ratio/low_mean': 0.06164874602109194, 'clip_ratio/low_min': 0.06164874602109194, 'clip_ratio/high_mean': 0.1036654431372881, 'clip_ratio/high_max': 0.1036654431372881, 'clip_ratio/region_mean': 0.16531418915838003, 'reward_total_mean': 0.2553885281085968, 'reward_meter_mean': 0.5915085673332214, 'reward_meter_std': 0.37578529119491577, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9946610331535339, 'reward_lexical_plausibility_std': 0.012002813629806042, 'reward_repeat_penalty_mean': 0.985454797744751, 'reward_repeat_penalty_std': 0.01338482741266489, 'reward_near_duplicate_penalty_mean': 0.9898806810379028, 'reward_near_duplicate_penalty_std': 0.004475778434425592, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9955277442932129, 'reward_distinct_2_std': 0.012649492360651493, 'reward_judge_quality_mean': 0.42499998211860657, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.2553885281085968, 'reward_total_composite_std': 0.16473935544490814, 'epoch': 0.01}

  8%|▊         | 262/3300 [58:44<10:04:31, 11.94s/it]INFO 04-12 11:44:57 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 263/3300 [58:55<9:36:10, 11.38s/it] 2026-04-12 11:45:07,064 | INFO | train_grpo_train | metrics_logged mode=train step=263

                                                    
{'loss': 0.0377, 'grad_norm': 23.47962760925293, 'learning_rate': 9.206060606060607e-06, 'num_tokens': 538274.0, 'completions/mean_length': 48.375, 'completions/min_length': 46.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.375, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.7191702127456665, 'rewards/meter/std': 0.3571155369281769, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.0353553369641304, 'rewards/repeat_penalty/mean': 0.97345370054245, 'rewards/repeat_penalty/std': 0.027753889560699463, 'rewards/near_duplicate_penalty/mean': 0.9798084497451782, 'rewards/near_duplicate_penalty/std': 0.011606940999627113, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9933686852455139, 'rewards/distinct_2/std': 0.01875614933669567, 'rewards/total_composite/mean': 0.24807783961296082, 'rewards/total_composite/std': 0.1624012291431427, 'reward': 0.24807783961296082, 'reward_std': 0.1624012291431427, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1425853669643402, 'sampling/sampling_logp_difference/max': 2.7772879600524902, 'sampling/importance_sampling_ratio/min': 0.06220698729157448, 'sampling/importance_sampling_ratio/mean': 0.9902973175048828, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5828147940337658, 'clip_ratio/low_mean': 0.04684942774474621, 'clip_ratio/low_min': 0.04684942774474621, 'clip_ratio/high_mean': 0.07390000717714429, 'clip_ratio/high_max': 0.07390000717714429, 'clip_ratio/region_mean': 0.1207494349218905, 'reward_total_mean': 0.24807783961296082, 'reward_meter_mean': 0.7191702127456665, 'reward_meter_std': 0.3571155369281769, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.97345370054245, 'reward_repeat_penalty_std': 0.027753889560699463, 'reward_near_duplicate_penalty_mean': 0.9798084497451782, 'reward_near_duplicate_penalty_std': 0.011606940999627113, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9933686852455139, 'reward_distinct_2_std': 0.01875614933669567, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.0353553369641304, 'reward_total_composite_mean': 0.24807783961296082, 'reward_total_composite_std': 0.1624012291431427, 'epoch': 0.01}

  8%|▊         | 263/3300 [58:55<9:36:10, 11.38s/it]INFO 04-12 11:45:07 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 264/3300 [59:05<9:20:57, 11.09s/it]2026-04-12 11:45:17,459 | INFO | train_grpo_train | metrics_logged mode=train step=264

                                                    
{'loss': 0.0832, 'grad_norm': 15.381331443786621, 'learning_rate': 9.203030303030304e-06, 'num_tokens': 539903.0, 'completions/mean_length': 44.625, 'completions/min_length': 37.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.625, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.48013317584991455, 'rewards/meter/std': 0.38423803448677063, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.05345224589109421, 'rewards/repeat_penalty/mean': 0.997526228427887, 'rewards/repeat_penalty/std': 0.004690530244261026, 'rewards/near_duplicate_penalty/mean': 0.997526228427887, 'rewards/near_duplicate_penalty/std': 0.004690530244261026, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.19194018840789795, 'rewards/total_composite/std': 0.15907663106918335, 'reward': 0.19194018840789795, 'reward_std': 0.15907663106918335, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1822473108768463, 'sampling/sampling_logp_difference/max': 1.3917350769042969, 'sampling/importance_sampling_ratio/min': 0.24864351749420166, 'sampling/importance_sampling_ratio/mean': 1.0437666177749634, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8692835420370102, 'clip_ratio/low_mean': 0.052580203395336866, 'clip_ratio/low_min': 0.052580203395336866, 'clip_ratio/high_mean': 0.08467308338731527, 'clip_ratio/high_max': 0.08467308338731527, 'clip_ratio/region_mean': 0.13725328678265214, 'reward_total_mean': 0.19194018840789795, 'reward_meter_mean': 0.48013317584991455, 'reward_meter_std': 0.38423803448677063, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.997526228427887, 'reward_repeat_penalty_std': 0.004690530244261026, 'reward_near_duplicate_penalty_mean': 0.997526228427887, 'reward_near_duplicate_penalty_std': 0.004690530244261026, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.05345224589109421, 'reward_total_composite_mean': 0.19194018840789795, 'reward_total_composite_std': 0.15907663106918335, 'epoch': 0.01}

  8%|▊         | 264/3300 [59:05<9:20:57, 11.09s/it]INFO 04-12 11:45:17 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 265/3300 [59:15<9:13:08, 10.94s/it]2026-04-12 11:45:28,042 | INFO | train_grpo_train | metrics_logged mode=train step=265

                                                    
{'loss': -0.0039, 'grad_norm': 9.906608581542969, 'learning_rate': 9.200000000000002e-06, 'num_tokens': 542069.0, 'completions/mean_length': 97.75, 'completions/min_length': 88.0, 'completions/max_length': 118.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 97.75, 'completions/min_terminated_length': 88.0, 'completions/max_terminated_length': 118.0, 'rewards/meter/mean': 0.8352808952331543, 'rewards/meter/std': 0.27949875593185425, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9836615324020386, 'rewards/lexical_plausibility/std': 0.046212125569581985, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.1060660183429718, 'rewards/repeat_penalty/mean': 0.9923805594444275, 'rewards/repeat_penalty/std': 0.005976299289613962, 'rewards/near_duplicate_penalty/mean': 0.9923805594444275, 'rewards/near_duplicate_penalty/std': 0.005976299289613962, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.30081719160079956, 'rewards/total_composite/std': 0.17533625662326813, 'reward': 0.30081719160079956, 'reward_std': 0.17533625662326813, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2217099666595459, 'sampling/sampling_logp_difference/max': 1.699728012084961, 'sampling/importance_sampling_ratio/min': 0.1827332228422165, 'sampling/importance_sampling_ratio/mean': 1.0484706163406372, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.641939789056778, 'clip_ratio/low_mean': 0.06033318303525448, 'clip_ratio/low_min': 0.06033318303525448, 'clip_ratio/high_mean': 0.12125750444829464, 'clip_ratio/high_max': 0.12125750444829464, 'clip_ratio/region_mean': 0.18159068748354912, 'reward_total_mean': 0.30081719160079956, 'reward_meter_mean': 0.8352808952331543, 'reward_meter_std': 0.27949875593185425, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.0707106739282608, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9836615324020386, 'reward_lexical_plausibility_std': 0.046212125569581985, 'reward_repeat_penalty_mean': 0.9923805594444275, 'reward_repeat_penalty_std': 0.005976299289613962, 'reward_near_duplicate_penalty_mean': 0.9923805594444275, 'reward_near_duplicate_penalty_std': 0.005976299289613962, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.1060660183429718, 'reward_total_composite_mean': 0.30081719160079956, 'reward_total_composite_std': 0.17533625662326813, 'epoch': 0.01}

  8%|▊         | 265/3300 [59:16<9:13:08, 10.94s/it]INFO 04-12 11:45:28 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 266/3300 [59:30<10:13:03, 12.12s/it]2026-04-12 11:45:42,939 | INFO | train_grpo_train | metrics_logged mode=train step=266

                                                     
{'loss': -0.0148, 'grad_norm': 5.331240653991699, 'learning_rate': 9.196969696969697e-06, 'num_tokens': 543768.0, 'completions/mean_length': 114.375, 'completions/min_length': 44.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 57.57143020629883, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.5551073551177979, 'rewards/meter/std': 0.4633782207965851, '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/lexical_plausibility/mean': 0.9571831226348877, 'rewards/lexical_plausibility/std': 0.12110446393489838, 'rewards/judge_quality/mean': 0.6137499809265137, 'rewards/judge_quality/std': 0.35419678688049316, 'rewards/repeat_penalty/mean': 0.9510653018951416, 'rewards/repeat_penalty/std': 0.07683956623077393, 'rewards/near_duplicate_penalty/mean': 0.9874784350395203, 'rewards/near_duplicate_penalty/std': 0.01600700430572033, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9826526641845703, 'rewards/distinct_2/std': 0.02612888813018799, 'rewards/total_composite/mean': 0.3820602595806122, 'rewards/total_composite/std': 0.36878228187561035, 'reward': 0.3820602595806122, 'reward_std': 0.36878228187561035, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18259519338607788, 'sampling/sampling_logp_difference/max': 1.6422357559204102, 'sampling/importance_sampling_ratio/min': 0.1935468316078186, 'sampling/importance_sampling_ratio/mean': 1.0315132141113281, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6832698434591293, 'clip_ratio/low_mean': 0.05894501693546772, 'clip_ratio/low_min': 0.05894501693546772, 'clip_ratio/high_mean': 0.106868801638484, 'clip_ratio/high_max': 0.106868801638484, 'clip_ratio/region_mean': 0.16581381857395172, 'reward_total_mean': 0.3820602595806122, 'reward_meter_mean': 0.5551073551177979, 'reward_meter_std': 0.4633782207965851, '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_lexical_plausibility_mean': 0.9571831226348877, 'reward_lexical_plausibility_std': 0.12110446393489838, 'reward_repeat_penalty_mean': 0.9510653018951416, 'reward_repeat_penalty_std': 0.07683956623077393, 'reward_near_duplicate_penalty_mean': 0.9874784350395203, 'reward_near_duplicate_penalty_std': 0.01600700430572033, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9826526641845703, 'reward_distinct_2_std': 0.02612888813018799, 'reward_judge_quality_mean': 0.6137499809265137, 'reward_judge_quality_std': 0.35419678688049316, 'reward_total_composite_mean': 0.3820602595806122, 'reward_total_composite_std': 0.36878228187561035, 'epoch': 0.01}

  8%|▊         | 266/3300 [59:30<10:13:03, 12.12s/it]INFO 04-12 11:45:43 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 267/3300 [59:40<9:32:18, 11.32s/it] 2026-04-12 11:45:52,390 | INFO | train_grpo_train | metrics_logged mode=train step=267

                                                    
{'loss': 0.0651, 'grad_norm': 17.208410263061523, 'learning_rate': 9.193939393939395e-06, 'num_tokens': 545491.0, 'completions/mean_length': 51.375, 'completions/min_length': 42.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.375, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.7053698301315308, 'rewards/meter/std': 0.29308009147644043, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.45875000953674316, 'rewards/judge_quality/std': 0.21390503644943237, 'rewards/repeat_penalty/mean': 0.9101802110671997, 'rewards/repeat_penalty/std': 0.09514571726322174, 'rewards/near_duplicate_penalty/mean': 0.9639840126037598, 'rewards/near_duplicate_penalty/std': 0.023830551654100418, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9425992369651794, 'rewards/distinct_2/std': 0.07865434885025024, 'rewards/total_composite/mean': 0.29284989833831787, 'rewards/total_composite/std': 0.1237848550081253, 'reward': 0.29284989833831787, 'reward_std': 0.1237848550081253, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2132163792848587, 'sampling/sampling_logp_difference/max': 2.403627395629883, 'sampling/importance_sampling_ratio/min': 0.09038948267698288, 'sampling/importance_sampling_ratio/mean': 1.025562047958374, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6964064985513687, 'clip_ratio/low_mean': 0.058897243812680244, 'clip_ratio/low_min': 0.058897243812680244, 'clip_ratio/high_mean': 0.11307246051728725, 'clip_ratio/high_max': 0.11307246051728725, 'clip_ratio/region_mean': 0.1719697043299675, 'reward_total_mean': 0.29284989833831787, 'reward_meter_mean': 0.7053698301315308, 'reward_meter_std': 0.29308009147644043, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9101802110671997, 'reward_repeat_penalty_std': 0.09514571726322174, 'reward_near_duplicate_penalty_mean': 0.9639840126037598, 'reward_near_duplicate_penalty_std': 0.023830551654100418, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9425992369651794, 'reward_distinct_2_std': 0.07865434885025024, 'reward_judge_quality_mean': 0.45875000953674316, 'reward_judge_quality_std': 0.21390503644943237, 'reward_total_composite_mean': 0.29284989833831787, 'reward_total_composite_std': 0.1237848550081253, 'epoch': 0.01}

  8%|▊         | 267/3300 [59:40<9:32:18, 11.32s/it]INFO 04-12 11:45:52 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 268/3300 [59:51<9:23:57, 11.16s/it]2026-04-12 11:46:03,173 | INFO | train_grpo_train | metrics_logged mode=train step=268

                                                    
{'loss': 0.0295, 'grad_norm': 13.87683391571045, 'learning_rate': 9.190909090909092e-06, 'num_tokens': 547163.0, 'completions/mean_length': 44.0, 'completions/min_length': 37.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.0, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.7814483046531677, 'rewards/meter/std': 0.31845104694366455, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9923516511917114, 'rewards/lexical_plausibility/std': 0.021632861346006393, 'rewards/judge_quality/mean': 0.6512500047683716, 'rewards/judge_quality/std': 0.29729434847831726, 'rewards/repeat_penalty/mean': 0.9803469777107239, 'rewards/repeat_penalty/std': 0.04630345478653908, 'rewards/near_duplicate_penalty/mean': 0.9893733263015747, 'rewards/near_duplicate_penalty/std': 0.02115357294678688, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.990384578704834, 'rewards/distinct_2/std': 0.027196412906050682, 'rewards/total_composite/mean': 0.5334149599075317, 'rewards/total_composite/std': 0.3403143286705017, 'reward': 0.5334149599075317, 'reward_std': 0.3403143584728241, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18669942021369934, 'sampling/sampling_logp_difference/max': 1.2455625534057617, 'sampling/importance_sampling_ratio/min': 0.28777897357940674, 'sampling/importance_sampling_ratio/mean': 1.0561602115631104, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4229229390621185, 'clip_ratio/low_mean': 0.10011645220220089, 'clip_ratio/low_min': 0.10011645220220089, 'clip_ratio/high_mean': 0.09112740121781826, 'clip_ratio/high_max': 0.09112740121781826, 'clip_ratio/region_mean': 0.19124385342001915, 'reward_total_mean': 0.5334149599075317, 'reward_meter_mean': 0.7814483046531677, 'reward_meter_std': 0.31845104694366455, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9923516511917114, 'reward_lexical_plausibility_std': 0.021632861346006393, 'reward_repeat_penalty_mean': 0.9803469777107239, 'reward_repeat_penalty_std': 0.04630345478653908, 'reward_near_duplicate_penalty_mean': 0.9893733263015747, 'reward_near_duplicate_penalty_std': 0.02115357294678688, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.990384578704834, 'reward_distinct_2_std': 0.027196412906050682, 'reward_judge_quality_mean': 0.6512500047683716, 'reward_judge_quality_std': 0.29729434847831726, 'reward_total_composite_mean': 0.5334149599075317, 'reward_total_composite_std': 0.3403143286705017, 'epoch': 0.01}

  8%|▊         | 268/3300 [59:51<9:23:57, 11.16s/it]INFO 04-12 11:46:03 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 269/3300 [1:00:02<9:26:51, 11.22s/it]2026-04-12 11:46:14,537 | INFO | train_grpo_train | metrics_logged mode=train step=269

                                                      
{'loss': 0.0004, 'grad_norm': 14.237889289855957, 'learning_rate': 9.187878787878789e-06, 'num_tokens': 549102.0, 'completions/mean_length': 55.375, 'completions/min_length': 48.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.375, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.5242666006088257, 'rewards/meter/std': 0.3175433874130249, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9946484565734863, 'rewards/lexical_plausibility/std': 0.015136471949517727, 'rewards/judge_quality/mean': 0.4087499976158142, 'rewards/judge_quality/std': 0.2290625423192978, 'rewards/repeat_penalty/mean': 0.9948704242706299, 'rewards/repeat_penalty/std': 0.006906564813107252, 'rewards/near_duplicate_penalty/mean': 0.9948704242706299, 'rewards/near_duplicate_penalty/std': 0.006906564813107252, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.24220727384090424, 'rewards/total_composite/std': 0.2698229253292084, 'reward': 0.24220727384090424, 'reward_std': 0.2698229253292084, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23273269832134247, 'sampling/sampling_logp_difference/max': 1.811131477355957, 'sampling/importance_sampling_ratio/min': 0.1634690761566162, 'sampling/importance_sampling_ratio/mean': 1.033265471458435, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2616512775421143, 'clip_ratio/low_mean': 0.1406971076503396, 'clip_ratio/low_min': 0.1406971076503396, 'clip_ratio/high_mean': 0.051724137738347054, 'clip_ratio/high_max': 0.051724137738347054, 'clip_ratio/region_mean': 0.19242124538868666, 'reward_total_mean': 0.24220727384090424, 'reward_meter_mean': 0.5242666006088257, 'reward_meter_std': 0.3175433874130249, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9946484565734863, 'reward_lexical_plausibility_std': 0.015136471949517727, 'reward_repeat_penalty_mean': 0.9948704242706299, 'reward_repeat_penalty_std': 0.006906564813107252, 'reward_near_duplicate_penalty_mean': 0.9948704242706299, 'reward_near_duplicate_penalty_std': 0.006906564813107252, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4087499976158142, 'reward_judge_quality_std': 0.2290625423192978, 'reward_total_composite_mean': 0.24220727384090424, 'reward_total_composite_std': 0.2698229253292084, 'epoch': 0.01}

  8%|▊         | 269/3300 [1:00:02<9:26:51, 11.22s/it]INFO 04-12 11:46:14 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 270/3300 [1:00:12<9:06:43, 10.83s/it]2026-04-12 11:46:24,512 | INFO | train_grpo_train | metrics_logged mode=train step=270

                                                      
{'loss': -0.0026, 'grad_norm': 16.44281578063965, 'learning_rate': 9.184848484848485e-06, 'num_tokens': 550687.0, 'completions/mean_length': 37.125, 'completions/min_length': 30.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.125, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.6374124884605408, 'rewards/meter/std': 0.41362813115119934, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.5299999713897705, 'rewards/judge_quality/std': 0.24512389302253723, 'rewards/repeat_penalty/mean': 0.9680814743041992, 'rewards/repeat_penalty/std': 0.038764070719480515, 'rewards/near_duplicate_penalty/mean': 0.9784848690032959, 'rewards/near_duplicate_penalty/std': 0.022901805117726326, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9893162250518799, 'rewards/distinct_2/std': 0.03021823801100254, 'rewards/total_composite/mean': 0.37814244627952576, 'rewards/total_composite/std': 0.3563656806945801, 'reward': 0.37814244627952576, 'reward_std': 0.3563656508922577, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2343471497297287, 'sampling/sampling_logp_difference/max': 2.7769792079925537, 'sampling/importance_sampling_ratio/min': 0.06222619488835335, 'sampling/importance_sampling_ratio/mean': 1.0300768613815308, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.157850459218025, 'clip_ratio/low_mean': 0.12842075154185295, 'clip_ratio/low_min': 0.12842075154185295, 'clip_ratio/high_mean': 0.06644042395055294, 'clip_ratio/high_max': 0.06644042395055294, 'clip_ratio/region_mean': 0.1948611754924059, 'reward_total_mean': 0.37814244627952576, 'reward_meter_mean': 0.6374124884605408, 'reward_meter_std': 0.41362813115119934, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.9680814743041992, 'reward_repeat_penalty_std': 0.038764070719480515, 'reward_near_duplicate_penalty_mean': 0.9784848690032959, 'reward_near_duplicate_penalty_std': 0.022901805117726326, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9893162250518799, 'reward_distinct_2_std': 0.03021823801100254, 'reward_judge_quality_mean': 0.5299999713897705, 'reward_judge_quality_std': 0.24512389302253723, 'reward_total_composite_mean': 0.37814244627952576, 'reward_total_composite_std': 0.3563656806945801, 'epoch': 0.01}

  8%|▊         | 270/3300 [1:00:12<9:06:43, 10.83s/it]INFO 04-12 11:46:24 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 271/3300 [1:00:24<9:20:08, 11.10s/it]2026-04-12 11:46:36,307 | INFO | train_grpo_train | metrics_logged mode=train step=271

                                                      
{'loss': -0.0001, 'grad_norm': 8.314339637756348, 'learning_rate': 9.181818181818184e-06, 'num_tokens': 553329.0, 'completions/mean_length': 148.25, 'completions/min_length': 130.0, 'completions/max_length': 164.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 148.25, 'completions/min_terminated_length': 130.0, 'completions/max_terminated_length': 164.0, 'rewards/meter/mean': 0.6872155666351318, 'rewards/meter/std': 0.23664095997810364, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.05345224589109421, 'rewards/repeat_penalty/mean': 0.9092499017715454, 'rewards/repeat_penalty/std': 0.05258341133594513, 'rewards/near_duplicate_penalty/mean': 0.9709391593933105, 'rewards/near_duplicate_penalty/std': 0.012902768328785896, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9420429468154907, 'rewards/distinct_2/std': 0.03494768589735031, 'rewards/total_composite/mean': 0.2294485718011856, 'rewards/total_composite/std': 0.07033482939004898, 'reward': 0.2294485718011856, 'reward_std': 0.07033482939004898, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21237437427043915, 'sampling/sampling_logp_difference/max': 1.4601354598999023, 'sampling/importance_sampling_ratio/min': 0.23220482468605042, 'sampling/importance_sampling_ratio/mean': 1.0503811836242676, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.7127927243709564, 'clip_ratio/low_mean': 0.08896524831652641, 'clip_ratio/low_min': 0.08896524831652641, 'clip_ratio/high_mean': 0.09131458215415478, 'clip_ratio/high_max': 0.09131458215415478, 'clip_ratio/region_mean': 0.1802798304706812, 'reward_total_mean': 0.2294485718011856, 'reward_meter_mean': 0.6872155666351318, 'reward_meter_std': 0.23664095997810364, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9092499017715454, 'reward_repeat_penalty_std': 0.05258341133594513, 'reward_near_duplicate_penalty_mean': 0.9709391593933105, 'reward_near_duplicate_penalty_std': 0.012902768328785896, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9420429468154907, 'reward_distinct_2_std': 0.03494768589735031, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.05345224589109421, 'reward_total_composite_mean': 0.2294485718011856, 'reward_total_composite_std': 0.07033482939004898, 'epoch': 0.01}

  8%|▊         | 271/3300 [1:00:24<9:20:08, 11.10s/it]INFO 04-12 11:46:36 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 272/3300 [1:00:39<10:26:37, 12.42s/it]2026-04-12 11:46:51,688 | INFO | train_grpo_train | metrics_logged mode=train step=272

                                                       
{'loss': -0.0299, 'grad_norm': 5.785467624664307, 'learning_rate': 9.178787878787879e-06, 'num_tokens': 555113.0, 'completions/mean_length': 117.0, 'completions/min_length': 54.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 60.57143020629883, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.5180156230926514, 'rewards/meter/std': 0.37577399611473083, '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/lexical_plausibility/mean': 0.9780136346817017, 'rewards/lexical_plausibility/std': 0.062186747789382935, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.1060660183429718, 'rewards/repeat_penalty/mean': 0.9848737716674805, 'rewards/repeat_penalty/std': 0.02308453619480133, 'rewards/near_duplicate_penalty/mean': 0.9916267991065979, 'rewards/near_duplicate_penalty/std': 0.006215302739292383, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9931318759918213, 'rewards/distinct_2/std': 0.01942601054906845, 'rewards/total_composite/mean': 0.1737125664949417, 'rewards/total_composite/std': 0.16292732954025269, 'reward': 0.1737125664949417, 'reward_std': 0.16292732954025269, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21715766191482544, 'sampling/sampling_logp_difference/max': 1.4731264114379883, 'sampling/importance_sampling_ratio/min': 0.22920776903629303, 'sampling/importance_sampling_ratio/mean': 1.0494691133499146, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.3031804263591766, 'clip_ratio/low_mean': 0.08485276810824871, 'clip_ratio/low_min': 0.08485276810824871, 'clip_ratio/high_mean': 0.07458556443452835, 'clip_ratio/high_max': 0.07458556443452835, 'clip_ratio/region_mean': 0.15943833254277706, 'reward_total_mean': 0.1737125664949417, 'reward_meter_mean': 0.5180156230926514, 'reward_meter_std': 0.37577399611473083, '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_lexical_plausibility_mean': 0.9780136346817017, 'reward_lexical_plausibility_std': 0.062186747789382935, 'reward_repeat_penalty_mean': 0.9848737716674805, 'reward_repeat_penalty_std': 0.02308453619480133, 'reward_near_duplicate_penalty_mean': 0.9916267991065979, 'reward_near_duplicate_penalty_std': 0.006215302739292383, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9931318759918213, 'reward_distinct_2_std': 0.01942601054906845, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.1060660183429718, 'reward_total_composite_mean': 0.1737125664949417, 'reward_total_composite_std': 0.16292732954025269, 'epoch': 0.01}

  8%|▊         | 272/3300 [1:00:39<10:26:37, 12.42s/it]INFO 04-12 11:46:51 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 273/3300 [1:00:50<9:58:05, 11.86s/it] 2026-04-12 11:47:02,293 | INFO | train_grpo_train | metrics_logged mode=train step=273

                                                      
{'loss': 0.015, 'grad_norm': 14.045431137084961, 'learning_rate': 9.175757575757576e-06, 'num_tokens': 556988.0, 'completions/mean_length': 68.375, 'completions/min_length': 62.0, 'completions/max_length': 82.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.375, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.7588969469070435, 'rewards/meter/std': 0.30399003624916077, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.14078858494758606, 'rewards/repeat_penalty/mean': 0.9785246849060059, 'rewards/repeat_penalty/std': 0.019320843741297722, 'rewards/near_duplicate_penalty/mean': 0.983896017074585, 'rewards/near_duplicate_penalty/std': 0.008063814602792263, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9945055246353149, 'rewards/distinct_2/std': 0.015540807507932186, 'rewards/total_composite/mean': 0.2500878572463989, 'rewards/total_composite/std': 0.18566662073135376, 'reward': 0.2500878572463989, 'reward_std': 0.18566662073135376, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22594274580478668, 'sampling/sampling_logp_difference/max': 1.8401718139648438, 'sampling/importance_sampling_ratio/min': 0.15879014134407043, 'sampling/importance_sampling_ratio/mean': 1.0287742614746094, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.28461155295372, 'clip_ratio/low_mean': 0.06879750732332468, 'clip_ratio/low_min': 0.06879750732332468, 'clip_ratio/high_mean': 0.12860195897519588, 'clip_ratio/high_max': 0.12860195897519588, 'clip_ratio/region_mean': 0.19739946629852057, 'reward_total_mean': 0.2500878572463989, 'reward_meter_mean': 0.7588969469070435, 'reward_meter_std': 0.30399003624916077, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9785246849060059, 'reward_repeat_penalty_std': 0.019320843741297722, 'reward_near_duplicate_penalty_mean': 0.983896017074585, 'reward_near_duplicate_penalty_std': 0.008063814602792263, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9945055246353149, 'reward_distinct_2_std': 0.015540807507932186, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.14078858494758606, 'reward_total_composite_mean': 0.2500878572463989, 'reward_total_composite_std': 0.18566662073135376, 'epoch': 0.01}

  8%|▊         | 273/3300 [1:00:50<9:58:05, 11.86s/it]INFO 04-12 11:47:02 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 274/3300 [1:00:59<9:21:29, 11.13s/it]2026-04-12 11:47:11,800 | INFO | train_grpo_train | metrics_logged mode=train step=274

                                                      
{'loss': 0.0594, 'grad_norm': 16.172096252441406, 'learning_rate': 9.172727272727274e-06, 'num_tokens': 558491.0, 'completions/mean_length': 33.875, 'completions/min_length': 29.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.875, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.16012103855609894, 'rewards/meter/std': 0.20767202973365784, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6137499809265137, 'rewards/judge_quality/std': 0.2558424770832062, 'rewards/repeat_penalty/mean': 0.9954314231872559, 'rewards/repeat_penalty/std': 0.006199869327247143, 'rewards/near_duplicate_penalty/mean': 0.9954314231872559, 'rewards/near_duplicate_penalty/std': 0.006199869327247143, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.10493992269039154, 'rewards/total_composite/std': 0.14338107407093048, 'reward': 0.10493992269039154, 'reward_std': 0.14338107407093048, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15259678661823273, 'sampling/sampling_logp_difference/max': 0.9395370483398438, 'sampling/importance_sampling_ratio/min': 0.39080873131752014, 'sampling/importance_sampling_ratio/mean': 1.0381031036376953, 'sampling/importance_sampling_ratio/max': 1.980162501335144, 'entropy': 1.3856611251831055, 'clip_ratio/low_mean': 0.09356303326785564, 'clip_ratio/low_min': 0.09356303326785564, 'clip_ratio/high_mean': 0.0356602817773819, 'clip_ratio/high_max': 0.0356602817773819, 'clip_ratio/region_mean': 0.12922331504523754, 'reward_total_mean': 0.10493992269039154, 'reward_meter_mean': 0.16012103855609894, 'reward_meter_std': 0.20767202973365784, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9954314231872559, 'reward_repeat_penalty_std': 0.006199869327247143, 'reward_near_duplicate_penalty_mean': 0.9954314231872559, 'reward_near_duplicate_penalty_std': 0.006199869327247143, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6137499809265137, 'reward_judge_quality_std': 0.2558424770832062, 'reward_total_composite_mean': 0.10493992269039154, 'reward_total_composite_std': 0.14338107407093048, 'epoch': 0.01}

  8%|▊         | 274/3300 [1:00:59<9:21:29, 11.13s/it]INFO 04-12 11:47:12 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 275/3300 [1:01:10<9:10:57, 10.93s/it]2026-04-12 11:47:22,110 | INFO | train_grpo_train | metrics_logged mode=train step=275

                                                      
{'loss': -0.0001, 'grad_norm': 24.670743942260742, 'learning_rate': 9.169696969696971e-06, 'num_tokens': 559879.0, 'completions/mean_length': 19.5, 'completions/min_length': 16.0, 'completions/max_length': 23.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.5, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.5803689956665039, 'rewards/meter/std': 0.4811611473560333, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.9275000095367432, 'rewards/judge_quality/std': 0.013887288980185986, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5340473651885986, 'rewards/total_composite/std': 0.44252634048461914, 'reward': 0.5340473651885986, 'reward_std': 0.44252628087997437, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1774034947156906, 'sampling/sampling_logp_difference/max': 1.35701322555542, 'sampling/importance_sampling_ratio/min': 0.25742852687835693, 'sampling/importance_sampling_ratio/mean': 0.9812389612197876, 'sampling/importance_sampling_ratio/max': 1.9948524236679077, 'entropy': 0.9840481281280518, 'clip_ratio/low_mean': 0.04843750037252903, 'clip_ratio/low_min': 0.04843750037252903, 'clip_ratio/high_mean': 0.10638116486370564, 'clip_ratio/high_max': 0.10638116486370564, 'clip_ratio/region_mean': 0.15481866523623466, 'reward_total_mean': 0.5340473651885986, 'reward_meter_mean': 0.5803689956665039, 'reward_meter_std': 0.4811611473560333, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.9275000095367432, 'reward_judge_quality_std': 0.013887288980185986, 'reward_total_composite_mean': 0.5340473651885986, 'reward_total_composite_std': 0.44252634048461914, 'epoch': 0.01}

  8%|▊         | 275/3300 [1:01:10<9:10:57, 10.93s/it]INFO 04-12 11:47:22 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 276/3300 [1:01:19<8:52:17, 10.56s/it]2026-04-12 11:47:31,814 | INFO | train_grpo_train | metrics_logged mode=train step=276

                                                      
{'loss': 0.0308, 'grad_norm': 16.561555862426758, 'learning_rate': 9.166666666666666e-06, 'num_tokens': 561403.0, 'completions/mean_length': 35.5, 'completions/min_length': 27.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.5, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.7875629663467407, 'rewards/meter/std': 0.3175351321697235, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9809887409210205, 'rewards/lexical_plausibility/std': 0.03529678285121918, 'rewards/judge_quality/mean': 0.5924999713897705, 'rewards/judge_quality/std': 0.25616681575775146, 'rewards/repeat_penalty/mean': 0.9985479712486267, 'rewards/repeat_penalty/std': 0.004106936510652304, 'rewards/near_duplicate_penalty/mean': 0.9985479712486267, 'rewards/near_duplicate_penalty/std': 0.004106936510652304, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.47072842717170715, 'rewards/total_composite/std': 0.26268860697746277, 'reward': 0.47072842717170715, 'reward_std': 0.26268860697746277, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2040463089942932, 'sampling/sampling_logp_difference/max': 1.1129522323608398, 'sampling/importance_sampling_ratio/min': 0.32858747243881226, 'sampling/importance_sampling_ratio/mean': 1.0465922355651855, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2738792449235916, 'clip_ratio/low_mean': 0.12270412687212229, 'clip_ratio/low_min': 0.12270412687212229, 'clip_ratio/high_mean': 0.03329741326160729, 'clip_ratio/high_max': 0.03329741326160729, 'clip_ratio/region_mean': 0.15600154013372958, 'reward_total_mean': 0.47072842717170715, 'reward_meter_mean': 0.7875629663467407, 'reward_meter_std': 0.3175351321697235, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9809887409210205, 'reward_lexical_plausibility_std': 0.03529678285121918, 'reward_repeat_penalty_mean': 0.9985479712486267, 'reward_repeat_penalty_std': 0.004106936510652304, 'reward_near_duplicate_penalty_mean': 0.9985479712486267, 'reward_near_duplicate_penalty_std': 0.004106936510652304, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5924999713897705, 'reward_judge_quality_std': 0.25616681575775146, 'reward_total_composite_mean': 0.47072842717170715, 'reward_total_composite_std': 0.26268860697746277, 'epoch': 0.01}

  8%|▊         | 276/3300 [1:01:19<8:52:17, 10.56s/it]INFO 04-12 11:47:32 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 277/3300 [1:01:29<8:34:23, 10.21s/it]2026-04-12 11:47:41,205 | INFO | train_grpo_train | metrics_logged mode=train step=277

                                                      
{'loss': 0.0254, 'grad_norm': 17.2047176361084, 'learning_rate': 9.163636363636365e-06, 'num_tokens': 562973.0, 'completions/mean_length': 38.25, 'completions/min_length': 32.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.25, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.6869758367538452, 'rewards/meter/std': 0.25142180919647217, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9735127091407776, 'rewards/lexical_plausibility/std': 0.050842247903347015, 'rewards/judge_quality/mean': 0.4674999713897705, 'rewards/judge_quality/std': 0.3022179901599884, 'rewards/repeat_penalty/mean': 0.9868708848953247, 'rewards/repeat_penalty/std': 0.013884245418012142, 'rewards/near_duplicate_penalty/mean': 0.9868708848953247, 'rewards/near_duplicate_penalty/std': 0.013884245418012142, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3224276900291443, 'rewards/total_composite/std': 0.19744513928890228, 'reward': 0.3224276900291443, 'reward_std': 0.1974451243877411, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21477508544921875, 'sampling/sampling_logp_difference/max': 1.3630951642990112, 'sampling/importance_sampling_ratio/min': 0.255867600440979, 'sampling/importance_sampling_ratio/mean': 1.020378828048706, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9873829185962677, 'clip_ratio/low_mean': 0.07072368264198303, 'clip_ratio/low_min': 0.07072368264198303, 'clip_ratio/high_mean': 0.12641051225364208, 'clip_ratio/high_max': 0.12641051225364208, 'clip_ratio/region_mean': 0.19713419489562511, 'reward_total_mean': 0.3224276900291443, 'reward_meter_mean': 0.6869758367538452, 'reward_meter_std': 0.25142180919647217, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9735127091407776, 'reward_lexical_plausibility_std': 0.050842247903347015, 'reward_repeat_penalty_mean': 0.9868708848953247, 'reward_repeat_penalty_std': 0.013884245418012142, 'reward_near_duplicate_penalty_mean': 0.9868708848953247, 'reward_near_duplicate_penalty_std': 0.013884245418012142, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4674999713897705, 'reward_judge_quality_std': 0.3022179901599884, 'reward_total_composite_mean': 0.3224276900291443, 'reward_total_composite_std': 0.19744513928890228, 'epoch': 0.01}

  8%|▊         | 277/3300 [1:01:29<8:34:23, 10.21s/it]INFO 04-12 11:47:41 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 278/3300 [1:01:43<9:40:49, 11.53s/it]2026-04-12 11:47:55,821 | INFO | train_grpo_train | metrics_logged mode=train step=278

                                                      
{'loss': -0.0731, 'grad_norm': 5.712137222290039, 'learning_rate': 9.160606060606061e-06, 'num_tokens': 565192.0, 'completions/mean_length': 129.375, 'completions/min_length': 70.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 74.71428680419922, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.42275065183639526, 'rewards/meter/std': 0.3581748306751251, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9331144094467163, 'rewards/lexical_plausibility/std': 0.18918108940124512, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.11649646610021591, 'rewards/repeat_penalty/mean': 0.9734548330307007, 'rewards/repeat_penalty/std': 0.02831222303211689, 'rewards/near_duplicate_penalty/mean': 0.9855933785438538, 'rewards/near_duplicate_penalty/std': 0.010175898671150208, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9876323938369751, 'rewards/distinct_2/std': 0.024720855057239532, 'rewards/total_composite/mean': 0.16008594632148743, 'rewards/total_composite/std': 0.14973433315753937, 'reward': 0.16008594632148743, 'reward_std': 0.14973431825637817, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2065655142068863, 'sampling/sampling_logp_difference/max': 1.6586599349975586, 'sampling/importance_sampling_ratio/min': 0.19039395451545715, 'sampling/importance_sampling_ratio/mean': 1.0514785051345825, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1411995738744736, 'clip_ratio/low_mean': 0.07053152285516262, 'clip_ratio/low_min': 0.07053152285516262, 'clip_ratio/high_mean': 0.10142524167895317, 'clip_ratio/high_max': 0.10142524167895317, 'clip_ratio/region_mean': 0.1719567645341158, 'reward_total_mean': 0.16008594632148743, 'reward_meter_mean': 0.42275065183639526, 'reward_meter_std': 0.3581748306751251, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.0707106739282608, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9331144094467163, 'reward_lexical_plausibility_std': 0.18918108940124512, 'reward_repeat_penalty_mean': 0.9734548330307007, 'reward_repeat_penalty_std': 0.02831222303211689, 'reward_near_duplicate_penalty_mean': 0.9855933785438538, 'reward_near_duplicate_penalty_std': 0.010175898671150208, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9876323938369751, 'reward_distinct_2_std': 0.024720855057239532, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.11649646610021591, 'reward_total_composite_mean': 0.16008594632148743, 'reward_total_composite_std': 0.14973433315753937, 'epoch': 0.01}

  8%|▊         | 278/3300 [1:01:43<9:40:49, 11.53s/it]INFO 04-12 11:47:56 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 279/3300 [1:01:53<9:11:42, 10.96s/it]2026-04-12 11:48:05,437 | INFO | train_grpo_train | metrics_logged mode=train step=279

                                                      
{'loss': 0.0602, 'grad_norm': 14.327668190002441, 'learning_rate': 9.157575757575758e-06, 'num_tokens': 566867.0, 'completions/mean_length': 39.375, 'completions/min_length': 36.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.375, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.4885554015636444, 'rewards/meter/std': 0.35893380641937256, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9716670513153076, 'rewards/lexical_plausibility/std': 0.04987131431698799, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.12817399203777313, 'rewards/repeat_penalty/mean': 0.9859256148338318, 'rewards/repeat_penalty/std': 0.015522838570177555, 'rewards/near_duplicate_penalty/mean': 0.9859256148338318, 'rewards/near_duplicate_penalty/std': 0.015522838570177555, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.149361714720726, 'rewards/total_composite/std': 0.1351163685321808, 'reward': 0.149361714720726, 'reward_std': 0.1351163536310196, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21950620412826538, 'sampling/sampling_logp_difference/max': 1.5779914855957031, 'sampling/importance_sampling_ratio/min': 0.20638921856880188, 'sampling/importance_sampling_ratio/mean': 1.0627535581588745, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.8971462845802307, 'clip_ratio/low_mean': 0.11658370168879628, 'clip_ratio/low_min': 0.11658370168879628, 'clip_ratio/high_mean': 0.048769925720989704, 'clip_ratio/high_max': 0.048769925720989704, 'clip_ratio/region_mean': 0.16535362740978599, 'reward_total_mean': 0.149361714720726, 'reward_meter_mean': 0.4885554015636444, 'reward_meter_std': 0.35893380641937256, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9716670513153076, 'reward_lexical_plausibility_std': 0.04987131431698799, 'reward_repeat_penalty_mean': 0.9859256148338318, 'reward_repeat_penalty_std': 0.015522838570177555, 'reward_near_duplicate_penalty_mean': 0.9859256148338318, 'reward_near_duplicate_penalty_std': 0.015522838570177555, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.12817399203777313, 'reward_total_composite_mean': 0.149361714720726, 'reward_total_composite_std': 0.1351163685321808, 'epoch': 0.01}

  8%|▊         | 279/3300 [1:01:53<9:11:42, 10.96s/it]INFO 04-12 11:48:05 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 280/3300 [1:02:08<10:10:01, 12.12s/it]2026-04-12 11:48:20,282 | INFO | train_grpo_train | metrics_logged mode=train step=280

                                                       
{'loss': -0.0629, 'grad_norm': 3.3239665031433105, 'learning_rate': 9.154545454545455e-06, 'num_tokens': 569347.0, 'completions/mean_length': 378.0, 'completions/min_length': 173.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 244.0, 'completions/min_terminated_length': 173.0, 'completions/max_terminated_length': 405.0, 'rewards/meter/mean': 0.2254655510187149, 'rewards/meter/std': 0.3117845058441162, 'rewards/count_adherence/mean': 0.7954545617103577, 'rewards/count_adherence/std': 0.2047257423400879, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/lexical_plausibility/mean': 0.8675926923751831, 'rewards/lexical_plausibility/std': 0.1524505615234375, 'rewards/judge_quality/mean': 0.125, 'rewards/judge_quality/std': 0.10350984334945679, 'rewards/repeat_penalty/mean': 0.9927418828010559, 'rewards/repeat_penalty/std': 0.0131080849096179, 'rewards/near_duplicate_penalty/mean': 0.9970248937606812, 'rewards/near_duplicate_penalty/std': 0.0018173862481489778, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.995694637298584, 'rewards/distinct_2/std': 0.012177488766610622, 'rewards/total_composite/mean': 0.04525914043188095, 'rewards/total_composite/std': 0.10281477868556976, 'reward': 0.04525914043188095, 'reward_std': 0.10281477868556976, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21117115020751953, 'sampling/sampling_logp_difference/max': 1.3579626083374023, 'sampling/importance_sampling_ratio/min': 0.257184237241745, 'sampling/importance_sampling_ratio/mean': 1.0500564575195312, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6696673035621643, 'clip_ratio/low_mean': 0.044871481135487556, 'clip_ratio/low_min': 0.044871481135487556, 'clip_ratio/high_mean': 0.023843931034207344, 'clip_ratio/high_max': 0.023843931034207344, 'clip_ratio/region_mean': 0.0687154121696949, 'reward_total_mean': 0.04525914043188095, 'reward_meter_mean': 0.2254655510187149, 'reward_meter_std': 0.3117845058441162, 'reward_count_adherence_mean': 0.7954545617103577, 'reward_count_adherence_std': 0.2047257423400879, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_lexical_plausibility_mean': 0.8675926923751831, 'reward_lexical_plausibility_std': 0.1524505615234375, 'reward_repeat_penalty_mean': 0.9927418828010559, 'reward_repeat_penalty_std': 0.0131080849096179, 'reward_near_duplicate_penalty_mean': 0.9970248937606812, 'reward_near_duplicate_penalty_std': 0.0018173862481489778, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.995694637298584, 'reward_distinct_2_std': 0.012177488766610622, 'reward_judge_quality_mean': 0.125, 'reward_judge_quality_std': 0.10350984334945679, 'reward_total_composite_mean': 0.04525914043188095, 'reward_total_composite_std': 0.10281477868556976, 'epoch': 0.01}

  8%|▊         | 280/3300 [1:02:08<10:10:01, 12.12s/it]INFO 04-12 11:48:20 [block_pool.py:378] Successfully reset prefix cache

  9%|▊         | 281/3300 [1:02:22<10:49:03, 12.90s/it]2026-04-12 11:48:34,994 | INFO | train_grpo_train | metrics_logged mode=train step=281

                                                       
{'loss': -0.082, 'grad_norm': 4.286274433135986, 'learning_rate': 9.151515151515153e-06, 'num_tokens': 571827.0, 'completions/mean_length': 240.0, 'completions/min_length': 137.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 149.33334350585938, 'completions/min_terminated_length': 137.0, 'completions/max_terminated_length': 169.0, 'rewards/meter/mean': 0.3047613501548767, 'rewards/meter/std': 0.18122979998588562, 'rewards/count_adherence/mean': 0.921875, 'rewards/count_adherence/std': 0.14846687018871307, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.8990951180458069, 'rewards/lexical_plausibility/std': 0.1883811503648758, 'rewards/judge_quality/mean': 0.2749999761581421, 'rewards/judge_quality/std': 0.16690459847450256, 'rewards/repeat_penalty/mean': 0.9925594329833984, 'rewards/repeat_penalty/std': 0.009130585938692093, 'rewards/near_duplicate_penalty/mean': 0.9951066374778748, 'rewards/near_duplicate_penalty/std': 0.004642870277166367, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9974359273910522, 'rewards/distinct_2/std': 0.007252376992255449, 'rewards/total_composite/mean': 0.07877501100301743, 'rewards/total_composite/std': 0.10420332103967667, 'reward': 0.07877501100301743, 'reward_std': 0.10420332103967667, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21506674587726593, 'sampling/sampling_logp_difference/max': 1.690638542175293, 'sampling/importance_sampling_ratio/min': 0.184401735663414, 'sampling/importance_sampling_ratio/mean': 1.0699589252471924, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.61429500579834, 'clip_ratio/low_mean': 0.050243156030774117, 'clip_ratio/low_min': 0.050243156030774117, 'clip_ratio/high_mean': 0.05922095663845539, 'clip_ratio/high_max': 0.05922095663845539, 'clip_ratio/region_mean': 0.10946411266922951, 'reward_total_mean': 0.07877501100301743, 'reward_meter_mean': 0.3047613501548767, 'reward_meter_std': 0.18122979998588562, 'reward_count_adherence_mean': 0.921875, 'reward_count_adherence_std': 0.14846687018871307, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.8990951180458069, 'reward_lexical_plausibility_std': 0.1883811503648758, 'reward_repeat_penalty_mean': 0.9925594329833984, 'reward_repeat_penalty_std': 0.009130585938692093, 'reward_near_duplicate_penalty_mean': 0.9951066374778748, 'reward_near_duplicate_penalty_std': 0.004642870277166367, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9974359273910522, 'reward_distinct_2_std': 0.007252376992255449, 'reward_judge_quality_mean': 0.2749999761581421, 'reward_judge_quality_std': 0.16690459847450256, 'reward_total_composite_mean': 0.07877501100301743, 'reward_total_composite_std': 0.10420332103967667, 'epoch': 0.01}

  9%|▊         | 281/3300 [1:02:22<10:49:03, 12.90s/it]INFO 04-12 11:48:35 [block_pool.py:378] Successfully reset prefix cache

  9%|▊         | 282/3300 [1:02:32<10:05:49, 12.04s/it]2026-04-12 11:48:45,039 | INFO | train_grpo_train | metrics_logged mode=train step=282

                                                       
{'loss': 0.0418, 'grad_norm': 14.189026832580566, 'learning_rate': 9.148484848484848e-06, 'num_tokens': 573417.0, 'completions/mean_length': 40.75, 'completions/min_length': 33.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.75, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.6186550855636597, 'rewards/meter/std': 0.3575173020362854, '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/lexical_plausibility/mean': 0.9828431606292725, 'rewards/lexical_plausibility/std': 0.048526935279369354, 'rewards/judge_quality/mean': 0.4000000059604645, 'rewards/judge_quality/std': 0.2507132589817047, 'rewards/repeat_penalty/mean': 0.9958095550537109, 'rewards/repeat_penalty/std': 0.00820265430957079, 'rewards/near_duplicate_penalty/mean': 0.9958095550537109, 'rewards/near_duplicate_penalty/std': 0.00820265430957079, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.28705093264579773, 'rewards/total_composite/std': 0.30884861946105957, 'reward': 0.28705093264579773, 'reward_std': 0.3088485896587372, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21589459478855133, 'sampling/sampling_logp_difference/max': 1.5428733825683594, 'sampling/importance_sampling_ratio/min': 0.21376599371433258, 'sampling/importance_sampling_ratio/mean': 1.046050786972046, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.9327259361743927, 'clip_ratio/low_mean': 0.08831252250820398, 'clip_ratio/low_min': 0.08831252250820398, 'clip_ratio/high_mean': 0.09627193212509155, 'clip_ratio/high_max': 0.09627193212509155, 'clip_ratio/region_mean': 0.18458445463329554, 'reward_total_mean': 0.28705093264579773, 'reward_meter_mean': 0.6186550855636597, 'reward_meter_std': 0.3575173020362854, '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_lexical_plausibility_mean': 0.9828431606292725, 'reward_lexical_plausibility_std': 0.048526935279369354, 'reward_repeat_penalty_mean': 0.9958095550537109, 'reward_repeat_penalty_std': 0.00820265430957079, 'reward_near_duplicate_penalty_mean': 0.9958095550537109, 'reward_near_duplicate_penalty_std': 0.00820265430957079, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4000000059604645, 'reward_judge_quality_std': 0.2507132589817047, 'reward_total_composite_mean': 0.28705093264579773, 'reward_total_composite_std': 0.30884861946105957, 'epoch': 0.01}

  9%|▊         | 282/3300 [1:02:33<10:05:49, 12.04s/it]INFO 04-12 11:48:45 [block_pool.py:378] Successfully reset prefix cache

  9%|▊         | 283/3300 [1:02:42<9:30:46, 11.35s/it] 2026-04-12 11:48:54,782 | INFO | train_grpo_train | metrics_logged mode=train step=283

                                                      
{'loss': 0.0456, 'grad_norm': 16.365924835205078, 'learning_rate': 9.145454545454546e-06, 'num_tokens': 574999.0, 'completions/mean_length': 34.75, 'completions/min_length': 29.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.75, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.5847836136817932, 'rewards/meter/std': 0.3397683799266815, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9473914504051208, 'rewards/lexical_plausibility/std': 0.11197065562009811, 'rewards/judge_quality/mean': 0.5674999952316284, 'rewards/judge_quality/std': 0.21756774187088013, 'rewards/repeat_penalty/mean': 0.9861968755722046, 'rewards/repeat_penalty/std': 0.0151191595941782, 'rewards/near_duplicate_penalty/mean': 0.9861968755722046, 'rewards/near_duplicate_penalty/std': 0.0151191595941782, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.33177411556243896, 'rewards/total_composite/std': 0.2660835385322571, 'reward': 0.33177411556243896, 'reward_std': 0.2660835385322571, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19750948250293732, 'sampling/sampling_logp_difference/max': 1.4663281440734863, 'sampling/importance_sampling_ratio/min': 0.2307712882757187, 'sampling/importance_sampling_ratio/mean': 1.0142182111740112, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8664885610342026, 'clip_ratio/low_mean': 0.11185344774276018, 'clip_ratio/low_min': 0.11185344774276018, 'clip_ratio/high_mean': 0.08955628052353859, 'clip_ratio/high_max': 0.08955628052353859, 'clip_ratio/region_mean': 0.20140972826629877, 'reward_total_mean': 0.33177411556243896, 'reward_meter_mean': 0.5847836136817932, 'reward_meter_std': 0.3397683799266815, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9473914504051208, 'reward_lexical_plausibility_std': 0.11197065562009811, 'reward_repeat_penalty_mean': 0.9861968755722046, 'reward_repeat_penalty_std': 0.0151191595941782, 'reward_near_duplicate_penalty_mean': 0.9861968755722046, 'reward_near_duplicate_penalty_std': 0.0151191595941782, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5674999952316284, 'reward_judge_quality_std': 0.21756774187088013, 'reward_total_composite_mean': 0.33177411556243896, 'reward_total_composite_std': 0.2660835385322571, 'epoch': 0.01}

  9%|▊         | 283/3300 [1:02:42<9:30:46, 11.35s/it]INFO 04-12 11:48:55 [block_pool.py:378] Successfully reset prefix cache

  9%|▊         | 284/3300 [1:02:52<9:01:21, 10.77s/it]2026-04-12 11:49:04,185 | INFO | train_grpo_train | metrics_logged mode=train step=284

                                                      
{'loss': -0.1131, 'grad_norm': 17.70599365234375, 'learning_rate': 9.142424242424243e-06, 'num_tokens': 576534.0, 'completions/mean_length': 37.875, 'completions/min_length': 17.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.875, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.8286744356155396, 'rewards/meter/std': 0.25950509309768677, '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/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.6349999904632568, 'rewards/judge_quality/std': 0.3184336721897125, 'rewards/repeat_penalty/mean': 0.9552540183067322, 'rewards/repeat_penalty/std': 0.08507763594388962, 'rewards/near_duplicate_penalty/mean': 0.9865040183067322, 'rewards/near_duplicate_penalty/std': 0.019740888848900795, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5359276533126831, 'rewards/total_composite/std': 0.3210074007511139, 'reward': 0.5359276533126831, 'reward_std': 0.3210074007511139, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19028252363204956, 'sampling/sampling_logp_difference/max': 1.4491653442382812, 'sampling/importance_sampling_ratio/min': 0.2347661703824997, 'sampling/importance_sampling_ratio/mean': 1.0386948585510254, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1328279227018356, 'clip_ratio/low_mean': 0.07423878833651543, 'clip_ratio/low_min': 0.07423878833651543, 'clip_ratio/high_mean': 0.07039508409798145, 'clip_ratio/high_max': 0.07039508409798145, 'clip_ratio/region_mean': 0.14463387243449688, 'reward_total_mean': 0.5359276533126831, 'reward_meter_mean': 0.8286744356155396, 'reward_meter_std': 0.25950509309768677, '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_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.9552540183067322, 'reward_repeat_penalty_std': 0.08507763594388962, 'reward_near_duplicate_penalty_mean': 0.9865040183067322, 'reward_near_duplicate_penalty_std': 0.019740888848900795, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6349999904632568, 'reward_judge_quality_std': 0.3184336721897125, 'reward_total_composite_mean': 0.5359276533126831, 'reward_total_composite_std': 0.3210074007511139, 'epoch': 0.01}

  9%|▊         | 284/3300 [1:02:52<9:01:21, 10.77s/it]INFO 04-12 11:49:04 [block_pool.py:378] Successfully reset prefix cache

  9%|▊         | 285/3300 [1:03:07<10:16:43, 12.27s/it]2026-04-12 11:49:19,974 | INFO | train_grpo_train | metrics_logged mode=train step=285

                                                       
{'loss': 0.0174, 'grad_norm': 6.271496295928955, 'learning_rate': 9.13939393939394e-06, 'num_tokens': 578739.0, 'completions/mean_length': 155.625, 'completions/min_length': 85.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 104.71428680419922, 'completions/min_terminated_length': 85.0, 'completions/max_terminated_length': 152.0, 'rewards/meter/mean': 0.5152211785316467, 'rewards/meter/std': 0.41637036204338074, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9600174427032471, 'rewards/lexical_plausibility/std': 0.11308781057596207, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.1414213478565216, 'rewards/repeat_penalty/mean': 0.9688326120376587, 'rewards/repeat_penalty/std': 0.0667465478181839, 'rewards/near_duplicate_penalty/mean': 0.986682653427124, 'rewards/near_duplicate_penalty/std': 0.0166612658649683, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9811463356018066, 'rewards/distinct_2/std': 0.05332630127668381, 'rewards/total_composite/mean': 0.17838908731937408, 'rewards/total_composite/std': 0.17776593565940857, 'reward': 0.17838908731937408, 'reward_std': 0.17776595056056976, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24316740036010742, 'sampling/sampling_logp_difference/max': 1.9065160751342773, 'sampling/importance_sampling_ratio/min': 0.14859719574451447, 'sampling/importance_sampling_ratio/mean': 1.0452635288238525, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.9215979874134064, 'clip_ratio/low_mean': 0.06691965274512768, 'clip_ratio/low_min': 0.06691965274512768, 'clip_ratio/high_mean': 0.1029907874763012, 'clip_ratio/high_max': 0.1029907874763012, 'clip_ratio/region_mean': 0.16991044022142887, 'reward_total_mean': 0.17838908731937408, 'reward_meter_mean': 0.5152211785316467, 'reward_meter_std': 0.41637036204338074, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9600174427032471, 'reward_lexical_plausibility_std': 0.11308781057596207, 'reward_repeat_penalty_mean': 0.9688326120376587, 'reward_repeat_penalty_std': 0.0667465478181839, 'reward_near_duplicate_penalty_mean': 0.986682653427124, 'reward_near_duplicate_penalty_std': 0.0166612658649683, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9811463356018066, 'reward_distinct_2_std': 0.05332630127668381, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.1414213478565216, 'reward_total_composite_mean': 0.17838908731937408, 'reward_total_composite_std': 0.17776593565940857, 'epoch': 0.01}

  9%|▊         | 285/3300 [1:03:07<10:16:43, 12.27s/it]INFO 04-12 11:49:20 [block_pool.py:378] Successfully reset prefix cache

  9%|▊         | 286/3300 [1:03:18<9:56:57, 11.88s/it] 2026-04-12 11:49:30,941 | INFO | train_grpo_train | metrics_logged mode=train step=286

                                                      
{'loss': 0.0581, 'grad_norm': 15.162145614624023, 'learning_rate': 9.136363636363637e-06, 'num_tokens': 580363.0, 'completions/mean_length': 42.0, 'completions/min_length': 38.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.0, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.48164448142051697, 'rewards/meter/std': 0.4154931604862213, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9786953330039978, 'rewards/lexical_plausibility/std': 0.06025872007012367, 'rewards/judge_quality/mean': 0.7137500047683716, 'rewards/judge_quality/std': 0.3153654634952545, 'rewards/repeat_penalty/mean': 0.9602888822555542, 'rewards/repeat_penalty/std': 0.03600199520587921, 'rewards/near_duplicate_penalty/mean': 0.9602888822555542, 'rewards/near_duplicate_penalty/std': 0.03600199520587921, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3208795189857483, 'rewards/total_composite/std': 0.33511683344841003, 'reward': 0.3208795189857483, 'reward_std': 0.33511683344841003, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1690378487110138, 'sampling/sampling_logp_difference/max': 1.516463279724121, 'sampling/importance_sampling_ratio/min': 0.2194867730140686, 'sampling/importance_sampling_ratio/mean': 1.0076239109039307, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1047745794057846, 'clip_ratio/low_mean': 0.09368273802101612, 'clip_ratio/low_min': 0.09368273802101612, 'clip_ratio/high_mean': 0.07388168945908546, 'clip_ratio/high_max': 0.07388168945908546, 'clip_ratio/region_mean': 0.16756442748010159, 'reward_total_mean': 0.3208795189857483, 'reward_meter_mean': 0.48164448142051697, 'reward_meter_std': 0.4154931604862213, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9786953330039978, 'reward_lexical_plausibility_std': 0.06025872007012367, 'reward_repeat_penalty_mean': 0.9602888822555542, 'reward_repeat_penalty_std': 0.03600199520587921, 'reward_near_duplicate_penalty_mean': 0.9602888822555542, 'reward_near_duplicate_penalty_std': 0.03600199520587921, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7137500047683716, 'reward_judge_quality_std': 0.3153654634952545, 'reward_total_composite_mean': 0.3208795189857483, 'reward_total_composite_std': 0.33511683344841003, 'epoch': 0.01}

  9%|▊         | 286/3300 [1:03:18<9:56:57, 11.88s/it]INFO 04-12 11:49:31 [block_pool.py:378] Successfully reset prefix cache

  9%|▊         | 287/3300 [1:03:29<9:33:41, 11.42s/it]2026-04-12 11:49:41,296 | INFO | train_grpo_train | metrics_logged mode=train step=287

                                                      
{'loss': -0.0182, 'grad_norm': 11.2236909866333, 'learning_rate': 9.133333333333335e-06, 'num_tokens': 582422.0, 'completions/mean_length': 72.375, 'completions/min_length': 66.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.375, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.9523226618766785, 'rewards/meter/std': 0.0559941865503788, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9989224076271057, 'rewards/lexical_plausibility/std': 0.0030478707049041986, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.9817270040512085, 'rewards/repeat_penalty/std': 0.038069162517786026, 'rewards/near_duplicate_penalty/mean': 0.9909766912460327, 'rewards/near_duplicate_penalty/std': 0.01231041457504034, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.990384578704834, 'rewards/distinct_2/std': 0.027196412906050682, 'rewards/total_composite/mean': 0.33471256494522095, 'rewards/total_composite/std': 0.09357573091983795, 'reward': 0.33471256494522095, 'reward_std': 0.09357573837041855, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22217997908592224, 'sampling/sampling_logp_difference/max': 1.4019737243652344, 'sampling/importance_sampling_ratio/min': 0.246110737323761, 'sampling/importance_sampling_ratio/mean': 1.0443822145462036, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.1826703250408173, 'clip_ratio/low_mean': 0.04497929662466049, 'clip_ratio/low_min': 0.04497929662466049, 'clip_ratio/high_mean': 0.14329812116920948, 'clip_ratio/high_max': 0.14329812116920948, 'clip_ratio/region_mean': 0.18827741779386997, 'reward_total_mean': 0.33471256494522095, 'reward_meter_mean': 0.9523226618766785, 'reward_meter_std': 0.0559941865503788, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9989224076271057, 'reward_lexical_plausibility_std': 0.0030478707049041986, 'reward_repeat_penalty_mean': 0.9817270040512085, 'reward_repeat_penalty_std': 0.038069162517786026, 'reward_near_duplicate_penalty_mean': 0.9909766912460327, 'reward_near_duplicate_penalty_std': 0.01231041457504034, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.990384578704834, 'reward_distinct_2_std': 0.027196412906050682, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.33471256494522095, 'reward_total_composite_std': 0.09357573091983795, 'epoch': 0.01}

  9%|▊         | 287/3300 [1:03:29<9:33:41, 11.42s/it]INFO 04-12 11:49:41 [block_pool.py:378] Successfully reset prefix cache

  9%|▊         | 288/3300 [1:03:38<9:02:33, 10.81s/it]2026-04-12 11:49:50,665 | INFO | train_grpo_train | metrics_logged mode=train step=288

                                                      
{'loss': -0.0021, 'grad_norm': 19.423803329467773, 'learning_rate': 9.130303030303032e-06, 'num_tokens': 583983.0, 'completions/mean_length': 33.125, 'completions/min_length': 29.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.125, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.9544972777366638, 'rewards/meter/std': 0.08394472301006317, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/repeat_penalty/mean': 0.9978380799293518, 'rewards/repeat_penalty/std': 0.004205068573355675, 'rewards/near_duplicate_penalty/mean': 0.9978380799293518, 'rewards/near_duplicate_penalty/std': 0.004205068573355675, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3095497488975525, 'rewards/total_composite/std': 0.07180063426494598, 'reward': 0.3095497488975525, 'reward_std': 0.07180062681436539, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10447724163532257, 'sampling/sampling_logp_difference/max': 2.5271964073181152, 'sampling/importance_sampling_ratio/min': 0.0798826664686203, 'sampling/importance_sampling_ratio/mean': 1.0091947317123413, 'sampling/importance_sampling_ratio/max': 1.7457364797592163, 'entropy': 0.4501609206199646, 'clip_ratio/low_mean': 0.02982292603701353, 'clip_ratio/low_min': 0.02982292603701353, 'clip_ratio/high_mean': 0.04191037290729582, 'clip_ratio/high_max': 0.04191037290729582, 'clip_ratio/region_mean': 0.07173329894430935, 'reward_total_mean': 0.3095497488975525, 'reward_meter_mean': 0.9544972777366638, 'reward_meter_std': 0.08394472301006317, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9978380799293518, 'reward_repeat_penalty_std': 0.004205068573355675, 'reward_near_duplicate_penalty_mean': 0.9978380799293518, 'reward_near_duplicate_penalty_std': 0.004205068573355675, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.3095497488975525, 'reward_total_composite_std': 0.07180063426494598, 'epoch': 0.01}

  9%|▊         | 288/3300 [1:03:38<9:02:33, 10.81s/it]INFO 04-12 11:49:50 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 289/3300 [1:03:48<8:47:40, 10.51s/it]2026-04-12 11:50:00,494 | INFO | train_grpo_train | metrics_logged mode=train step=289

                                                      
{'loss': 0.0271, 'grad_norm': 17.322498321533203, 'learning_rate': 9.127272727272727e-06, 'num_tokens': 585554.0, 'completions/mean_length': 35.375, 'completions/min_length': 27.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.375, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.7997351884841919, 'rewards/meter/std': 0.3051901161670685, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5012500286102295, 'rewards/judge_quality/std': 0.36286312341690063, 'rewards/repeat_penalty/mean': 0.9882843494415283, 'rewards/repeat_penalty/std': 0.015018452890217304, 'rewards/near_duplicate_penalty/mean': 0.9882843494415283, 'rewards/near_duplicate_penalty/std': 0.015018452890217304, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4522401690483093, 'rewards/total_composite/std': 0.37616705894470215, 'reward': 0.4522401690483093, 'reward_std': 0.37616702914237976, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20104411244392395, 'sampling/sampling_logp_difference/max': 1.4658613204956055, 'sampling/importance_sampling_ratio/min': 0.2308790385723114, 'sampling/importance_sampling_ratio/mean': 1.013503074645996, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4743967205286026, 'clip_ratio/low_mean': 0.11661775130778551, 'clip_ratio/low_min': 0.11661775130778551, 'clip_ratio/high_mean': 0.07630914449691772, 'clip_ratio/high_max': 0.07630914449691772, 'clip_ratio/region_mean': 0.19292689580470324, 'reward_total_mean': 0.4522401690483093, 'reward_meter_mean': 0.7997351884841919, 'reward_meter_std': 0.3051901161670685, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9882843494415283, 'reward_repeat_penalty_std': 0.015018452890217304, 'reward_near_duplicate_penalty_mean': 0.9882843494415283, 'reward_near_duplicate_penalty_std': 0.015018452890217304, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5012500286102295, 'reward_judge_quality_std': 0.36286312341690063, 'reward_total_composite_mean': 0.4522401690483093, 'reward_total_composite_std': 0.37616705894470215, 'epoch': 0.01}

  9%|▉         | 289/3300 [1:03:48<8:47:40, 10.51s/it]INFO 04-12 11:50:00 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 290/3300 [1:03:59<8:56:22, 10.69s/it]2026-04-12 11:50:11,602 | INFO | train_grpo_train | metrics_logged mode=train step=290

                                                      
{'loss': 0.0015, 'grad_norm': 15.765957832336426, 'learning_rate': 9.124242424242425e-06, 'num_tokens': 587475.0, 'completions/mean_length': 57.125, 'completions/min_length': 54.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.125, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.6953085660934448, 'rewards/meter/std': 0.38268086314201355, '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/lexical_plausibility/mean': 0.9712885022163391, 'rewards/lexical_plausibility/std': 0.08120834082365036, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.9866654276847839, 'rewards/repeat_penalty/std': 0.027135929092764854, 'rewards/near_duplicate_penalty/mean': 0.9919589161872864, 'rewards/near_duplicate_penalty/std': 0.012913516722619534, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9945055246353149, 'rewards/distinct_2/std': 0.015540807507932186, 'rewards/total_composite/mean': 0.20763850212097168, 'rewards/total_composite/std': 0.18367838859558105, 'reward': 0.20763850212097168, 'reward_std': 0.18367838859558105, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22305156290531158, 'sampling/sampling_logp_difference/max': 1.2211408615112305, 'sampling/importance_sampling_ratio/min': 0.2948935329914093, 'sampling/importance_sampling_ratio/mean': 1.0478311777114868, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.9298156797885895, 'clip_ratio/low_mean': 0.09555976465344429, 'clip_ratio/low_min': 0.09555976465344429, 'clip_ratio/high_mean': 0.10091807879507542, 'clip_ratio/high_max': 0.10091807879507542, 'clip_ratio/region_mean': 0.1964778434485197, 'reward_total_mean': 0.20763850212097168, 'reward_meter_mean': 0.6953085660934448, 'reward_meter_std': 0.38268086314201355, '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_lexical_plausibility_mean': 0.9712885022163391, 'reward_lexical_plausibility_std': 0.08120834082365036, 'reward_repeat_penalty_mean': 0.9866654276847839, 'reward_repeat_penalty_std': 0.027135929092764854, 'reward_near_duplicate_penalty_mean': 0.9919589161872864, 'reward_near_duplicate_penalty_std': 0.012913516722619534, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9945055246353149, 'reward_distinct_2_std': 0.015540807507932186, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.20763850212097168, 'reward_total_composite_std': 0.18367838859558105, 'epoch': 0.01}

  9%|▉         | 290/3300 [1:03:59<8:56:22, 10.69s/it]INFO 04-12 11:50:11 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 291/3300 [1:04:10<8:55:38, 10.68s/it]2026-04-12 11:50:22,257 | INFO | train_grpo_train | metrics_logged mode=train step=291

                                                      
{'loss': 0.0553, 'grad_norm': 25.532514572143555, 'learning_rate': 9.121212121212122e-06, 'num_tokens': 589009.0, 'completions/mean_length': 36.75, 'completions/min_length': 33.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.75, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.6063839197158813, 'rewards/meter/std': 0.4344141185283661, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7100000381469727, 'rewards/judge_quality/std': 0.3123185336589813, 'rewards/repeat_penalty/mean': 0.9811705350875854, 'rewards/repeat_penalty/std': 0.044920410960912704, 'rewards/near_duplicate_penalty/mean': 0.9946225881576538, 'rewards/near_duplicate_penalty/std': 0.009966662153601646, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9862637519836426, 'rewards/distinct_2/std': 0.0388520210981369, 'rewards/total_composite/mean': 0.43089666962623596, 'rewards/total_composite/std': 0.3954765498638153, 'reward': 0.43089666962623596, 'reward_std': 0.3954765498638153, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2248818725347519, 'sampling/sampling_logp_difference/max': 2.0057125091552734, 'sampling/importance_sampling_ratio/min': 0.1345643848180771, 'sampling/importance_sampling_ratio/mean': 0.9709168076515198, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2025741189718246, 'clip_ratio/low_mean': 0.11072261445224285, 'clip_ratio/low_min': 0.11072261445224285, 'clip_ratio/high_mean': 0.08179810643196106, 'clip_ratio/high_max': 0.08179810643196106, 'clip_ratio/region_mean': 0.1925207208842039, 'reward_total_mean': 0.43089666962623596, 'reward_meter_mean': 0.6063839197158813, 'reward_meter_std': 0.4344141185283661, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9811705350875854, 'reward_repeat_penalty_std': 0.044920410960912704, 'reward_near_duplicate_penalty_mean': 0.9946225881576538, 'reward_near_duplicate_penalty_std': 0.009966662153601646, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9862637519836426, 'reward_distinct_2_std': 0.0388520210981369, 'reward_judge_quality_mean': 0.7100000381469727, 'reward_judge_quality_std': 0.3123185336589813, 'reward_total_composite_mean': 0.43089666962623596, 'reward_total_composite_std': 0.3954765498638153, 'epoch': 0.01}

  9%|▉         | 291/3300 [1:04:10<8:55:38, 10.68s/it]INFO 04-12 11:50:22 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 292/3300 [1:04:21<9:05:46, 10.89s/it]2026-04-12 11:50:33,622 | INFO | train_grpo_train | metrics_logged mode=train step=292

                                                      
{'loss': 0.012, 'grad_norm': 8.342798233032227, 'learning_rate': 9.118181818181819e-06, 'num_tokens': 591660.0, 'completions/mean_length': 117.375, 'completions/min_length': 109.0, 'completions/max_length': 136.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 117.375, 'completions/min_terminated_length': 109.0, 'completions/max_terminated_length': 136.0, 'rewards/meter/mean': 0.8820005059242249, 'rewards/meter/std': 0.24240130186080933, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/repeat_penalty/mean': 0.9748523235321045, 'rewards/repeat_penalty/std': 0.033441174775362015, 'rewards/near_duplicate_penalty/mean': 0.9887571334838867, 'rewards/near_duplicate_penalty/std': 0.01039181649684906, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9857282638549805, 'rewards/distinct_2/std': 0.024325896054506302, 'rewards/total_composite/mean': 0.24187153577804565, 'rewards/total_composite/std': 0.13933728635311127, 'reward': 0.24187153577804565, 'reward_std': 0.13933728635311127, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22239527106285095, 'sampling/sampling_logp_difference/max': 1.9674361944198608, 'sampling/importance_sampling_ratio/min': 0.1398148536682129, 'sampling/importance_sampling_ratio/mean': 1.0495963096618652, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.499325841665268, 'clip_ratio/low_mean': 0.04573502764105797, 'clip_ratio/low_min': 0.04573502764105797, 'clip_ratio/high_mean': 0.11654209718108177, 'clip_ratio/high_max': 0.11654209718108177, 'clip_ratio/region_mean': 0.16227712482213974, 'reward_total_mean': 0.24187153577804565, 'reward_meter_mean': 0.8820005059242249, 'reward_meter_std': 0.24240130186080933, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9748523235321045, 'reward_repeat_penalty_std': 0.033441174775362015, 'reward_near_duplicate_penalty_mean': 0.9887571334838867, 'reward_near_duplicate_penalty_std': 0.01039181649684906, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9857282638549805, 'reward_distinct_2_std': 0.024325896054506302, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.24187153577804565, 'reward_total_composite_std': 0.13933728635311127, 'epoch': 0.01}

  9%|▉         | 292/3300 [1:04:21<9:05:46, 10.89s/it]INFO 04-12 11:50:33 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 293/3300 [1:04:36<10:02:03, 12.01s/it]2026-04-12 11:50:48,266 | INFO | train_grpo_train | metrics_logged mode=train step=293

                                                       
{'loss': -0.0571, 'grad_norm': 2.7340917587280273, 'learning_rate': 9.115151515151516e-06, 'num_tokens': 593283.0, 'completions/mean_length': 161.875, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 45.16666793823242, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.77762371301651, 'rewards/meter/std': 0.3516334593296051, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.8469035625457764, 'rewards/lexical_plausibility/std': 0.23500525951385498, 'rewards/judge_quality/mean': 0.5637500286102295, 'rewards/judge_quality/std': 0.3871484696865082, 'rewards/repeat_penalty/mean': 0.994547426700592, 'rewards/repeat_penalty/std': 0.008622530847787857, 'rewards/near_duplicate_penalty/mean': 0.994547426700592, 'rewards/near_duplicate_penalty/std': 0.008622530847787857, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4651144742965698, 'rewards/total_composite/std': 0.39441993832588196, 'reward': 0.4651144742965698, 'reward_std': 0.39441990852355957, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20372723042964935, 'sampling/sampling_logp_difference/max': 1.4653916358947754, 'sampling/importance_sampling_ratio/min': 0.2309875190258026, 'sampling/importance_sampling_ratio/mean': 1.0500022172927856, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5260865688323975, 'clip_ratio/low_mean': 0.050378479063510895, 'clip_ratio/low_min': 0.050378479063510895, 'clip_ratio/high_mean': 0.06681599840521812, 'clip_ratio/high_max': 0.06681599840521812, 'clip_ratio/region_mean': 0.11719447746872902, 'reward_total_mean': 0.4651144742965698, 'reward_meter_mean': 0.77762371301651, 'reward_meter_std': 0.3516334593296051, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.8469035625457764, 'reward_lexical_plausibility_std': 0.23500525951385498, 'reward_repeat_penalty_mean': 0.994547426700592, 'reward_repeat_penalty_std': 0.008622530847787857, 'reward_near_duplicate_penalty_mean': 0.994547426700592, 'reward_near_duplicate_penalty_std': 0.008622530847787857, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5637500286102295, 'reward_judge_quality_std': 0.3871484696865082, 'reward_total_composite_mean': 0.4651144742965698, 'reward_total_composite_std': 0.39441993832588196, 'epoch': 0.01}

  9%|▉         | 293/3300 [1:04:36<10:02:03, 12.01s/it]INFO 04-12 11:50:48 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 294/3300 [1:04:51<10:57:03, 13.12s/it]2026-04-12 11:51:03,952 | INFO | train_grpo_train | metrics_logged mode=train step=294

                                                       
{'loss': -0.0628, 'grad_norm': 5.737674713134766, 'learning_rate': 9.112121212121214e-06, 'num_tokens': 594913.0, 'completions/mean_length': 97.75, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 38.57143020629883, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.5892068147659302, 'rewards/meter/std': 0.4092927575111389, '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/lexical_plausibility/mean': 0.861744225025177, 'rewards/lexical_plausibility/std': 0.2673322260379791, 'rewards/judge_quality/mean': 0.45499998331069946, 'rewards/judge_quality/std': 0.31883716583251953, 'rewards/repeat_penalty/mean': 0.968208909034729, 'rewards/repeat_penalty/std': 0.0881827101111412, 'rewards/near_duplicate_penalty/mean': 0.999458909034729, 'rewards/near_duplicate_penalty/std': 0.0015305416891351342, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2709839642047882, 'rewards/total_composite/std': 0.2503100633621216, 'reward': 0.2709839642047882, 'reward_std': 0.2503100335597992, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1945735514163971, 'sampling/sampling_logp_difference/max': 1.0185413360595703, 'sampling/importance_sampling_ratio/min': 0.3611213266849518, 'sampling/importance_sampling_ratio/mean': 1.0450493097305298, 'sampling/importance_sampling_ratio/max': 1.8643429279327393, 'entropy': 1.9573396891355515, 'clip_ratio/low_mean': 0.0683183167129755, 'clip_ratio/low_min': 0.0683183167129755, 'clip_ratio/high_mean': 0.08630233816802502, 'clip_ratio/high_max': 0.08630233816802502, 'clip_ratio/region_mean': 0.15462065488100052, 'reward_total_mean': 0.2709839642047882, 'reward_meter_mean': 0.5892068147659302, 'reward_meter_std': 0.4092927575111389, '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_lexical_plausibility_mean': 0.861744225025177, 'reward_lexical_plausibility_std': 0.2673322260379791, 'reward_repeat_penalty_mean': 0.968208909034729, 'reward_repeat_penalty_std': 0.0881827101111412, 'reward_near_duplicate_penalty_mean': 0.999458909034729, 'reward_near_duplicate_penalty_std': 0.0015305416891351342, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.45499998331069946, 'reward_judge_quality_std': 0.31883716583251953, 'reward_total_composite_mean': 0.2709839642047882, 'reward_total_composite_std': 0.2503100633621216, 'epoch': 0.01}

  9%|▉         | 294/3300 [1:04:51<10:57:03, 13.12s/it]INFO 04-12 11:51:04 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 295/3300 [1:05:05<11:08:06, 13.34s/it]2026-04-12 11:51:17,815 | INFO | train_grpo_train | metrics_logged mode=train step=295

                                                       
{'loss': -0.0126, 'grad_norm': 5.329678535461426, 'learning_rate': 9.10909090909091e-06, 'num_tokens': 596280.0, 'completions/mean_length': 80.875, 'completions/min_length': 15.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 19.285715103149414, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.7221282720565796, 'rewards/meter/std': 0.4450470209121704, '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/lexical_plausibility/mean': 0.903509259223938, 'rewards/lexical_plausibility/std': 0.16369616985321045, 'rewards/judge_quality/mean': 0.5475000143051147, 'rewards/judge_quality/std': 0.4066500961780548, 'rewards/repeat_penalty/mean': 0.7480332255363464, 'rewards/repeat_penalty/std': 0.005562899634242058, 'rewards/near_duplicate_penalty/mean': 0.9973776340484619, 'rewards/near_duplicate_penalty/std': 0.007417213171720505, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4099704623222351, 'rewards/total_composite/std': 0.4066827595233917, 'reward': 0.4099704623222351, 'reward_std': 0.4066827595233917, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19901323318481445, 'sampling/sampling_logp_difference/max': 0.7967100143432617, 'sampling/importance_sampling_ratio/min': 0.45080968737602234, 'sampling/importance_sampling_ratio/mean': 1.0342069864273071, 'sampling/importance_sampling_ratio/max': 1.895004391670227, 'entropy': 2.027714654803276, 'clip_ratio/low_mean': 0.08847841061651707, 'clip_ratio/low_min': 0.08847841061651707, 'clip_ratio/high_mean': 0.06477591255679727, 'clip_ratio/high_max': 0.06477591255679727, 'clip_ratio/region_mean': 0.15325432317331433, 'reward_total_mean': 0.4099704623222351, 'reward_meter_mean': 0.7221282720565796, 'reward_meter_std': 0.4450470209121704, '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_lexical_plausibility_mean': 0.903509259223938, 'reward_lexical_plausibility_std': 0.16369616985321045, 'reward_repeat_penalty_mean': 0.7480332255363464, 'reward_repeat_penalty_std': 0.005562899634242058, 'reward_near_duplicate_penalty_mean': 0.9973776340484619, 'reward_near_duplicate_penalty_std': 0.007417213171720505, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5475000143051147, 'reward_judge_quality_std': 0.4066500961780548, 'reward_total_composite_mean': 0.4099704623222351, 'reward_total_composite_std': 0.4066827595233917, 'epoch': 0.01}

  9%|▉         | 295/3300 [1:05:05<11:08:06, 13.34s/it]INFO 04-12 11:51:18 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 296/3300 [1:05:14<9:53:32, 11.85s/it] 2026-04-12 11:51:26,203 | INFO | train_grpo_train | metrics_logged mode=train step=296

                                                      
{'loss': 0.0344, 'grad_norm': 21.586929321289062, 'learning_rate': 9.106060606060606e-06, 'num_tokens': 597767.0, 'completions/mean_length': 23.875, 'completions/min_length': 19.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.875, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.4130640923976898, 'rewards/meter/std': 0.4726019501686096, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.984375, 'rewards/lexical_plausibility/std': 0.0289318785071373, 'rewards/judge_quality/mean': 0.8424999713897705, 'rewards/judge_quality/std': 0.28014031052589417, 'rewards/repeat_penalty/mean': 0.7448863983154297, 'rewards/repeat_penalty/std': 0.014463555067777634, 'rewards/near_duplicate_penalty/mean': 0.9931818246841431, 'rewards/near_duplicate_penalty/std': 0.01928473263978958, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.29227522015571594, 'rewards/total_composite/std': 0.39396294951438904, 'reward': 0.29227522015571594, 'reward_std': 0.39396294951438904, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21043173968791962, 'sampling/sampling_logp_difference/max': 1.9541807174682617, 'sampling/importance_sampling_ratio/min': 0.1416804939508438, 'sampling/importance_sampling_ratio/mean': 1.0021095275878906, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8546439111232758, 'clip_ratio/low_mean': 0.12120436877012253, 'clip_ratio/low_min': 0.12120436877012253, 'clip_ratio/high_mean': 0.030869564972817898, 'clip_ratio/high_max': 0.030869564972817898, 'clip_ratio/region_mean': 0.15207393374294043, 'reward_total_mean': 0.29227522015571594, 'reward_meter_mean': 0.4130640923976898, 'reward_meter_std': 0.4726019501686096, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.984375, 'reward_lexical_plausibility_std': 0.0289318785071373, 'reward_repeat_penalty_mean': 0.7448863983154297, 'reward_repeat_penalty_std': 0.014463555067777634, 'reward_near_duplicate_penalty_mean': 0.9931818246841431, 'reward_near_duplicate_penalty_std': 0.01928473263978958, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8424999713897705, 'reward_judge_quality_std': 0.28014031052589417, 'reward_total_composite_mean': 0.29227522015571594, 'reward_total_composite_std': 0.39396294951438904, 'epoch': 0.01}

  9%|▉         | 296/3300 [1:05:14<9:53:32, 11.85s/it]INFO 04-12 11:51:26 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 297/3300 [1:05:23<9:14:23, 11.08s/it]2026-04-12 11:51:35,470 | INFO | train_grpo_train | metrics_logged mode=train step=297

                                                      
{'loss': -0.0373, 'grad_norm': 15.426389694213867, 'learning_rate': 9.103030303030304e-06, 'num_tokens': 599328.0, 'completions/mean_length': 38.125, 'completions/min_length': 34.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.125, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.76540607213974, 'rewards/meter/std': 0.23013561964035034, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.44624999165534973, 'rewards/judge_quality/std': 0.2173829823732376, 'rewards/repeat_penalty/mean': 0.9947516322135925, 'rewards/repeat_penalty/std': 0.008241748437285423, 'rewards/near_duplicate_penalty/mean': 0.9947516322135925, 'rewards/near_duplicate_penalty/std': 0.008241748437285423, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.34635573625564575, 'rewards/total_composite/std': 0.23247207701206207, 'reward': 0.34635573625564575, 'reward_std': 0.23247207701206207, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2140551656484604, 'sampling/sampling_logp_difference/max': 1.916731357574463, 'sampling/importance_sampling_ratio/min': 0.1470869481563568, 'sampling/importance_sampling_ratio/mean': 1.0367729663848877, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0692486613988876, 'clip_ratio/low_mean': 0.12777711264789104, 'clip_ratio/low_min': 0.12777711264789104, 'clip_ratio/high_mean': 0.06627468205988407, 'clip_ratio/high_max': 0.06627468205988407, 'clip_ratio/region_mean': 0.19405179470777512, 'reward_total_mean': 0.34635573625564575, 'reward_meter_mean': 0.76540607213974, 'reward_meter_std': 0.23013561964035034, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9947516322135925, 'reward_repeat_penalty_std': 0.008241748437285423, 'reward_near_duplicate_penalty_mean': 0.9947516322135925, 'reward_near_duplicate_penalty_std': 0.008241748437285423, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.44624999165534973, 'reward_judge_quality_std': 0.2173829823732376, 'reward_total_composite_mean': 0.34635573625564575, 'reward_total_composite_std': 0.23247207701206207, 'epoch': 0.01}

  9%|▉         | 297/3300 [1:05:23<9:14:23, 11.08s/it]INFO 04-12 11:51:35 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 298/3300 [1:05:32<8:44:17, 10.48s/it]2026-04-12 11:51:44,550 | INFO | train_grpo_train | metrics_logged mode=train step=298

                                                      
{'loss': -0.0198, 'grad_norm': 15.232839584350586, 'learning_rate': 9.100000000000001e-06, 'num_tokens': 600785.0, 'completions/mean_length': 33.125, 'completions/min_length': 28.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.125, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.3792986273765564, 'rewards/meter/std': 0.30770623683929443, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9464869499206543, 'rewards/lexical_plausibility/std': 0.15135782957077026, 'rewards/judge_quality/mean': 0.6637499928474426, 'rewards/judge_quality/std': 0.28465205430984497, 'rewards/repeat_penalty/mean': 0.9954763650894165, 'rewards/repeat_penalty/std': 0.006409325636923313, 'rewards/near_duplicate_penalty/mean': 0.9954763650894165, 'rewards/near_duplicate_penalty/std': 0.006409325636923313, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.26781579852104187, 'rewards/total_composite/std': 0.2779422402381897, 'reward': 0.26781579852104187, 'reward_std': 0.2779422402381897, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23376497626304626, 'sampling/sampling_logp_difference/max': 1.7966036796569824, 'sampling/importance_sampling_ratio/min': 0.16586124897003174, 'sampling/importance_sampling_ratio/mean': 0.9879671931266785, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7423544377088547, 'clip_ratio/low_mean': 0.07093166094273329, 'clip_ratio/low_min': 0.07093166094273329, 'clip_ratio/high_mean': 0.08968173898756504, 'clip_ratio/high_max': 0.08968173898756504, 'clip_ratio/region_mean': 0.16061339993029833, 'reward_total_mean': 0.26781579852104187, 'reward_meter_mean': 0.3792986273765564, 'reward_meter_std': 0.30770623683929443, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9464869499206543, 'reward_lexical_plausibility_std': 0.15135782957077026, 'reward_repeat_penalty_mean': 0.9954763650894165, 'reward_repeat_penalty_std': 0.006409325636923313, 'reward_near_duplicate_penalty_mean': 0.9954763650894165, 'reward_near_duplicate_penalty_std': 0.006409325636923313, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6637499928474426, 'reward_judge_quality_std': 0.28465205430984497, 'reward_total_composite_mean': 0.26781579852104187, 'reward_total_composite_std': 0.2779422402381897, 'epoch': 0.01}

  9%|▉         | 298/3300 [1:05:32<8:44:17, 10.48s/it]INFO 04-12 11:51:44 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 299/3300 [1:05:42<8:35:16, 10.30s/it]2026-04-12 11:51:54,441 | INFO | train_grpo_train | metrics_logged mode=train step=299

                                                      
{'loss': -0.0217, 'grad_norm': 12.565971374511719, 'learning_rate': 9.096969696969698e-06, 'num_tokens': 602675.0, 'completions/mean_length': 65.25, 'completions/min_length': 56.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.25, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.3651045262813568, 'rewards/meter/std': 0.26593083143234253, '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/lexical_plausibility/mean': 0.9940476417541504, 'rewards/lexical_plausibility/std': 0.016835875809192657, 'rewards/judge_quality/mean': 0.4624999761581421, 'rewards/judge_quality/std': 0.12464234977960587, 'rewards/repeat_penalty/mean': 0.9899188876152039, 'rewards/repeat_penalty/std': 0.017242584377527237, 'rewards/near_duplicate_penalty/mean': 0.9952132701873779, 'rewards/near_duplicate_penalty/std': 0.004840857349336147, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9946581125259399, 'rewards/distinct_2/std': 0.01510911900550127, 'rewards/total_composite/mean': 0.16748185455799103, 'rewards/total_composite/std': 0.1259952336549759, 'reward': 0.16748185455799103, 'reward_std': 0.1259952336549759, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2156853824853897, 'sampling/sampling_logp_difference/max': 1.4517087936401367, 'sampling/importance_sampling_ratio/min': 0.23416979610919952, 'sampling/importance_sampling_ratio/mean': 1.0100153684616089, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.418099120259285, 'clip_ratio/low_mean': 0.0630058366805315, 'clip_ratio/low_min': 0.0630058366805315, 'clip_ratio/high_mean': 0.1380710806697607, 'clip_ratio/high_max': 0.1380710806697607, 'clip_ratio/region_mean': 0.2010769173502922, 'reward_total_mean': 0.16748185455799103, 'reward_meter_mean': 0.3651045262813568, 'reward_meter_std': 0.26593083143234253, '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_lexical_plausibility_mean': 0.9940476417541504, 'reward_lexical_plausibility_std': 0.016835875809192657, 'reward_repeat_penalty_mean': 0.9899188876152039, 'reward_repeat_penalty_std': 0.017242584377527237, 'reward_near_duplicate_penalty_mean': 0.9952132701873779, 'reward_near_duplicate_penalty_std': 0.004840857349336147, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9946581125259399, 'reward_distinct_2_std': 0.01510911900550127, 'reward_judge_quality_mean': 0.4624999761581421, 'reward_judge_quality_std': 0.12464234977960587, 'reward_total_composite_mean': 0.16748185455799103, 'reward_total_composite_std': 0.1259952336549759, 'epoch': 0.01}

  9%|▉         | 299/3300 [1:05:42<8:35:16, 10.30s/it]INFO 04-12 11:51:54 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 300/3300 [1:05:57<9:49:43, 11.79s/it]2026-04-12 11:52:09,716 | INFO | train_grpo_train | metrics_logged mode=train step=300

                                                      
{'loss': -0.0387, 'grad_norm': 5.960353374481201, 'learning_rate': 9.093939393939395e-06, 'num_tokens': 604355.0, 'completions/mean_length': 112.0, 'completions/min_length': 48.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 54.857147216796875, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.6109622716903687, 'rewards/meter/std': 0.3946131765842438, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9230944514274597, 'rewards/lexical_plausibility/std': 0.21752171218395233, 'rewards/judge_quality/mean': 0.4087499976158142, 'rewards/judge_quality/std': 0.3223324716091156, 'rewards/repeat_penalty/mean': 0.9986394643783569, 'rewards/repeat_penalty/std': 0.0025408314540982246, 'rewards/near_duplicate_penalty/mean': 0.9986394643783569, 'rewards/near_duplicate_penalty/std': 0.0025408314540982246, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.27971965074539185, 'rewards/total_composite/std': 0.27698031067848206, 'reward': 0.27971965074539185, 'reward_std': 0.27698028087615967, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21483492851257324, 'sampling/sampling_logp_difference/max': 1.8838939666748047, 'sampling/importance_sampling_ratio/min': 0.15199707448482513, 'sampling/importance_sampling_ratio/mean': 1.0404177904129028, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.3810570687055588, 'clip_ratio/low_mean': 0.061070866882801056, 'clip_ratio/low_min': 0.061070866882801056, 'clip_ratio/high_mean': 0.09348621964454651, 'clip_ratio/high_max': 0.09348621964454651, 'clip_ratio/region_mean': 0.15455708652734756, 'reward_total_mean': 0.27971965074539185, 'reward_meter_mean': 0.6109622716903687, 'reward_meter_std': 0.3946131765842438, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9230944514274597, 'reward_lexical_plausibility_std': 0.21752171218395233, 'reward_repeat_penalty_mean': 0.9986394643783569, 'reward_repeat_penalty_std': 0.0025408314540982246, 'reward_near_duplicate_penalty_mean': 0.9986394643783569, 'reward_near_duplicate_penalty_std': 0.0025408314540982246, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4087499976158142, 'reward_judge_quality_std': 0.3223324716091156, 'reward_total_composite_mean': 0.27971965074539185, 'reward_total_composite_std': 0.27698031067848206, 'epoch': 0.01}

  9%|▉         | 300/3300 [1:05:57<9:49:43, 11.79s/it]INFO 04-12 11:52:09 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


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

 15%|█▌        | 2/13 [00:09<00:54,  4.92s/it]

 23%|██▎       | 3/13 [00:16<00:58,  5.81s/it]

 31%|███       | 4/13 [00:28<01:11,  7.91s/it]

 38%|███▊      | 5/13 [00:36<01:03,  7.97s/it]

 46%|████▌     | 6/13 [00:45<00:59,  8.50s/it]

 54%|█████▍    | 7/13 [00:58<00:58,  9.68s/it]

 62%|██████▏   | 8/13 [01:10<00:52, 10.46s/it]

 69%|██████▉   | 9/13 [01:20<00:41, 10.37s/it]

 77%|███████▋  | 10/13 [01:33<00:33, 11.08s/it]

 85%|████████▍ | 11/13 [01:41<00:20, 10.37s/it]

 92%|█████████▏| 12/13 [01:50<00:09,  9.72s/it]

100%|██████████| 13/13 [01:58<00:00,  9.25s/it]2026-04-12 11:54:20,753 | INFO | train_grpo_train | metrics_logged mode=eval step=300

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 131.0085, 'eval_samples_per_second': 0.794, 'eval_steps_per_second': 0.099, 'eval_num_tokens': 604355.0, 'eval_completions/mean_length': 147.91346153846155, 'eval_completions/min_length': 32.92307692307692, 'eval_completions/max_length': 354.0, 'eval_completions/clipped_ratio': 0.0673076923076923, 'eval_completions/mean_terminated_length': 121.22445267897386, 'eval_completions/min_terminated_length': 32.92307692307692, 'eval_completions/max_terminated_length': 257.53846153846155, 'eval_rewards/meter/mean': 0.45110395092230576, 'eval_rewards/meter/std': 0.38209012150764465, 'eval_rewards/count_adherence/mean': 0.9236263220126812, 'eval_rewards/count_adherence/std': 0.12292832809572037, 'eval_rewards/arabic_clean/mean': 0.8365384615384616, 'eval_rewards/arabic_clean/std': 0.2923306066256303, 'eval_rewards/lexical_plausibility/mean': 0.9571945438018212, 'eval_rewards/lexical_plausibility/std': 0.09503344768801561, 'eval_rewards/judge_quality/mean': 0.365769230402433, 'eval_rewards/judge_quality/std': 0.2133815953364739, 'eval_rewards/repeat_penalty/mean': 0.9364584180024954, 'eval_rewards/repeat_penalty/std': 0.10855107381939888, 'eval_rewards/near_duplicate_penalty/mean': 0.9848691683549148, 'eval_rewards/near_duplicate_penalty/std': 0.019361631479114294, 'eval_rewards/opening_diversity/mean': 0.9926382211538461, 'eval_rewards/opening_diversity/std': 0.02082225461848653, 'eval_rewards/distinct_2/mean': 0.9587259521851172, 'eval_rewards/distinct_2/std': 0.08783535759609479, 'eval_rewards/total_composite/mean': 0.15561715857340738, 'eval_rewards/total_composite/std': 0.1566895954310894, 'eval_reward': 0.15561715857340738, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.15180421677919534, 'eval_sampling/sampling_logp_difference/max': 1.1537190950833833, 'eval_sampling/importance_sampling_ratio/min': 0.31711981617487395, 'eval_sampling/importance_sampling_ratio/mean': 1.0492328772178063, 'eval_sampling/importance_sampling_ratio/max': 1.7163617610931396, 'eval_entropy': 2.512108170069181, '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': 0.15561715857340738, 'eval_reward_meter_mean': 0.45110395092230576, 'eval_reward_meter_std': 0.38209012150764465, 'eval_reward_count_adherence_mean': 0.9236263220126812, 'eval_reward_count_adherence_std': 0.12292832809572037, 'eval_reward_arabic_clean_mean': 0.8365384615384616, 'eval_reward_arabic_clean_std': 0.2923306066256303, 'eval_reward_lexical_plausibility_mean': 0.9571945438018212, 'eval_reward_lexical_plausibility_std': 0.09503344768801561, 'eval_reward_repeat_penalty_mean': 0.9364584180024954, 'eval_reward_repeat_penalty_std': 0.10855107381939888, 'eval_reward_near_duplicate_penalty_mean': 0.9848691683549148, 'eval_reward_near_duplicate_penalty_std': 0.019361631479114294, 'eval_reward_opening_diversity_mean': 0.9926382211538461, 'eval_reward_opening_diversity_std': 0.02082225461848653, 'eval_reward_distinct_2_mean': 0.9587259521851172, 'eval_reward_distinct_2_std': 0.08783535759609479, 'eval_reward_judge_quality_mean': 0.365769230402433, 'eval_reward_judge_quality_std': 0.2133815953364739, 'eval_reward_total_composite_mean': 0.15561715857340738, 'eval_reward_total_composite_std': 0.1566895954310894, 'epoch': 0.01}

  9%|▉         | 300/3300 [1:08:08<9:49:43, 11.79s/it]

100%|██████████| 13/13 [01:58<00:00,  9.25s/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(

  9%|▉         | 301/3300 [1:08:20<42:41:39, 51.25s/it]2026-04-12 11:54:33,033 | INFO | train_grpo_train | metrics_logged mode=train step=301

                                                       
{'loss': 0.0362, 'grad_norm': 15.899238586425781, 'learning_rate': 9.090909090909091e-06, 'num_tokens': 605887.0, 'completions/mean_length': 36.5, 'completions/min_length': 28.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.5, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.5783378481864929, 'rewards/meter/std': 0.4569213092327118, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.47874999046325684, 'rewards/judge_quality/std': 0.1683056354522705, 'rewards/repeat_penalty/mean': 0.9830946922302246, 'rewards/repeat_penalty/std': 0.02091960422694683, 'rewards/near_duplicate_penalty/mean': 0.9830946922302246, 'rewards/near_duplicate_penalty/std': 0.02091960422694683, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2994975745677948, 'rewards/total_composite/std': 0.27936550974845886, 'reward': 0.2994975745677948, 'reward_std': 0.2793654799461365, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1651681810617447, 'sampling/sampling_logp_difference/max': 0.998927116394043, 'sampling/importance_sampling_ratio/min': 0.36827436089515686, 'sampling/importance_sampling_ratio/mean': 1.0356861352920532, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0033613443374634, 'clip_ratio/low_mean': 0.09417487867176533, 'clip_ratio/low_min': 0.09417487867176533, 'clip_ratio/high_mean': 0.0902948398143053, 'clip_ratio/high_max': 0.0902948398143053, 'clip_ratio/region_mean': 0.18446971848607063, 'reward_total_mean': 0.2994975745677948, 'reward_meter_mean': 0.5783378481864929, 'reward_meter_std': 0.4569213092327118, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9830946922302246, 'reward_repeat_penalty_std': 0.02091960422694683, 'reward_near_duplicate_penalty_mean': 0.9830946922302246, 'reward_near_duplicate_penalty_std': 0.02091960422694683, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.47874999046325684, 'reward_judge_quality_std': 0.1683056354522705, 'reward_total_composite_mean': 0.2994975745677948, 'reward_total_composite_std': 0.27936550974845886, 'epoch': 0.01}

  9%|▉         | 301/3300 [1:08:21<42:41:39, 51.25s/it]INFO 04-12 11:54:33 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 302/3300 [1:08:30<32:14:21, 38.71s/it]2026-04-12 11:54:42,489 | INFO | train_grpo_train | metrics_logged mode=train step=302

                                                       
{'loss': -0.0167, 'grad_norm': 15.26990032196045, 'learning_rate': 9.087878787878788e-06, 'num_tokens': 607692.0, 'completions/mean_length': 55.625, 'completions/min_length': 47.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.625, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.7832177877426147, 'rewards/meter/std': 0.3151888847351074, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9977678656578064, 'rewards/lexical_plausibility/std': 0.006313450634479523, 'rewards/judge_quality/mean': 0.41624999046325684, 'rewards/judge_quality/std': 0.20900700986385345, 'rewards/repeat_penalty/mean': 0.938164234161377, 'rewards/repeat_penalty/std': 0.07573452591896057, 'rewards/near_duplicate_penalty/mean': 0.958517849445343, 'rewards/near_duplicate_penalty/std': 0.040154967457056046, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9774928689002991, 'rewards/distinct_2/std': 0.04526444524526596, 'rewards/total_composite/mean': 0.36066925525665283, 'rewards/total_composite/std': 0.24422086775302887, 'reward': 0.36066925525665283, 'reward_std': 0.24422085285186768, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2266959547996521, 'sampling/sampling_logp_difference/max': 1.6513595581054688, 'sampling/importance_sampling_ratio/min': 0.19178898632526398, 'sampling/importance_sampling_ratio/mean': 1.0588935613632202, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1657568216323853, 'clip_ratio/low_mean': 0.13119650445878506, 'clip_ratio/low_min': 0.13119650445878506, 'clip_ratio/high_mean': 0.06655014585703611, 'clip_ratio/high_max': 0.06655014585703611, 'clip_ratio/region_mean': 0.19774665031582117, 'reward_total_mean': 0.36066925525665283, 'reward_meter_mean': 0.7832177877426147, 'reward_meter_std': 0.3151888847351074, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9977678656578064, 'reward_lexical_plausibility_std': 0.006313450634479523, 'reward_repeat_penalty_mean': 0.938164234161377, 'reward_repeat_penalty_std': 0.07573452591896057, 'reward_near_duplicate_penalty_mean': 0.958517849445343, 'reward_near_duplicate_penalty_std': 0.040154967457056046, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9774928689002991, 'reward_distinct_2_std': 0.04526444524526596, 'reward_judge_quality_mean': 0.41624999046325684, 'reward_judge_quality_std': 0.20900700986385345, 'reward_total_composite_mean': 0.36066925525665283, 'reward_total_composite_std': 0.24422086775302887, 'epoch': 0.01}

  9%|▉         | 302/3300 [1:08:30<32:14:21, 38.71s/it]INFO 04-12 11:54:42 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 303/3300 [1:08:44<26:00:32, 31.24s/it]2026-04-12 11:54:56,300 | INFO | train_grpo_train | metrics_logged mode=train step=303

                                                       
{'loss': 0.0061, 'grad_norm': 4.482475280761719, 'learning_rate': 9.084848484848486e-06, 'num_tokens': 609013.0, 'completions/mean_length': 85.125, 'completions/min_length': 18.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 24.142858505249023, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.5166142582893372, 'rewards/meter/std': 0.5072239637374878, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9274661540985107, 'rewards/lexical_plausibility/std': 0.13499270379543304, 'rewards/judge_quality/mean': 0.4300000071525574, 'rewards/judge_quality/std': 0.32262319326400757, 'rewards/repeat_penalty/mean': 0.78125, 'rewards/repeat_penalty/std': 0.0883883461356163, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.27508029341697693, 'rewards/total_composite/std': 0.40899473428726196, 'reward': 0.27508029341697693, 'reward_std': 0.40899476408958435, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2104078233242035, 'sampling/sampling_logp_difference/max': 1.2070770263671875, 'sampling/importance_sampling_ratio/min': 0.29907017946243286, 'sampling/importance_sampling_ratio/mean': 1.0715430974960327, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.986671194434166, 'clip_ratio/low_mean': 0.11303419061005116, 'clip_ratio/low_min': 0.11303419061005116, 'clip_ratio/high_mean': 0.07534373831003904, 'clip_ratio/high_max': 0.07534373831003904, 'clip_ratio/region_mean': 0.1883779289200902, 'reward_total_mean': 0.27508029341697693, 'reward_meter_mean': 0.5166142582893372, 'reward_meter_std': 0.5072239637374878, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9274661540985107, 'reward_lexical_plausibility_std': 0.13499270379543304, 'reward_repeat_penalty_mean': 0.78125, 'reward_repeat_penalty_std': 0.0883883461356163, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4300000071525574, 'reward_judge_quality_std': 0.32262319326400757, 'reward_total_composite_mean': 0.27508029341697693, 'reward_total_composite_std': 0.40899473428726196, 'epoch': 0.01}

  9%|▉         | 303/3300 [1:08:44<26:00:32, 31.24s/it]INFO 04-12 11:54:56 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 304/3300 [1:08:54<20:47:57, 24.99s/it]2026-04-12 11:55:06,711 | INFO | train_grpo_train | metrics_logged mode=train step=304

                                                       
{'loss': 0.0822, 'grad_norm': 8.455896377563477, 'learning_rate': 9.081818181818183e-06, 'num_tokens': 611509.0, 'completions/mean_length': 134.0, 'completions/min_length': 114.0, 'completions/max_length': 149.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 134.0, 'completions/min_terminated_length': 114.0, 'completions/max_terminated_length': 149.0, 'rewards/meter/mean': 0.5805405378341675, 'rewards/meter/std': 0.2993761897087097, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.998106062412262, 'rewards/lexical_plausibility/std': 0.005356864538043737, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.8322775363922119, 'rewards/repeat_penalty/std': 0.15659010410308838, 'rewards/near_duplicate_penalty/mean': 0.9583142399787903, 'rewards/near_duplicate_penalty/std': 0.030610177665948868, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.881321132183075, 'rewards/distinct_2/std': 0.10168742388486862, 'rewards/total_composite/mean': 0.21507972478866577, 'rewards/total_composite/std': 0.1228046640753746, 'reward': 0.21507972478866577, 'reward_std': 0.122804656624794, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20804409682750702, 'sampling/sampling_logp_difference/max': 1.6914610862731934, 'sampling/importance_sampling_ratio/min': 0.1842501312494278, 'sampling/importance_sampling_ratio/mean': 1.0556303262710571, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4510658532381058, 'clip_ratio/low_mean': 0.06107951141893864, 'clip_ratio/low_min': 0.06107951141893864, 'clip_ratio/high_mean': 0.1182973887771368, 'clip_ratio/high_max': 0.1182973887771368, 'clip_ratio/region_mean': 0.17937690019607544, 'reward_total_mean': 0.21507972478866577, 'reward_meter_mean': 0.5805405378341675, 'reward_meter_std': 0.2993761897087097, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.998106062412262, 'reward_lexical_plausibility_std': 0.005356864538043737, 'reward_repeat_penalty_mean': 0.8322775363922119, 'reward_repeat_penalty_std': 0.15659010410308838, 'reward_near_duplicate_penalty_mean': 0.9583142399787903, 'reward_near_duplicate_penalty_std': 0.030610177665948868, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.881321132183075, 'reward_distinct_2_std': 0.10168742388486862, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.21507972478866577, 'reward_total_composite_std': 0.1228046640753746, 'epoch': 0.01}

  9%|▉         | 304/3300 [1:08:54<20:47:57, 24.99s/it]INFO 04-12 11:55:06 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 305/3300 [1:09:04<17:02:59, 20.49s/it]2026-04-12 11:55:16,707 | INFO | train_grpo_train | metrics_logged mode=train step=305

                                                       
{'loss': 0.0182, 'grad_norm': 9.742971420288086, 'learning_rate': 9.078787878787878e-06, 'num_tokens': 613414.0, 'completions/mean_length': 73.125, 'completions/min_length': 68.0, 'completions/max_length': 82.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.125, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.39772510528564453, 'rewards/meter/std': 0.2838525176048279, '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/lexical_plausibility/mean': 0.9150548577308655, 'rewards/lexical_plausibility/std': 0.10696487873792648, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.11649646610021591, 'rewards/repeat_penalty/mean': 0.9957312345504761, 'rewards/repeat_penalty/std': 0.0034778863191604614, 'rewards/near_duplicate_penalty/mean': 0.9957312345504761, 'rewards/near_duplicate_penalty/std': 0.0034778863191604614, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.12142910063266754, 'rewards/total_composite/std': 0.10667760670185089, 'reward': 0.12142910063266754, 'reward_std': 0.10667760670185089, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22550351917743683, 'sampling/sampling_logp_difference/max': 1.464259147644043, 'sampling/importance_sampling_ratio/min': 0.23124924302101135, 'sampling/importance_sampling_ratio/mean': 1.054365873336792, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.6886763870716095, 'clip_ratio/low_mean': 0.0966362003237009, 'clip_ratio/low_min': 0.0966362003237009, 'clip_ratio/high_mean': 0.08360382542014122, 'clip_ratio/high_max': 0.08360382542014122, 'clip_ratio/region_mean': 0.18024002574384212, 'reward_total_mean': 0.12142910063266754, 'reward_meter_mean': 0.39772510528564453, 'reward_meter_std': 0.2838525176048279, '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_lexical_plausibility_mean': 0.9150548577308655, 'reward_lexical_plausibility_std': 0.10696487873792648, 'reward_repeat_penalty_mean': 0.9957312345504761, 'reward_repeat_penalty_std': 0.0034778863191604614, 'reward_near_duplicate_penalty_mean': 0.9957312345504761, 'reward_near_duplicate_penalty_std': 0.0034778863191604614, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.11649646610021591, 'reward_total_composite_mean': 0.12142910063266754, 'reward_total_composite_std': 0.10667760670185089, 'epoch': 0.01}

  9%|▉         | 305/3300 [1:09:04<17:02:59, 20.49s/it]INFO 04-12 11:55:16 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 306/3300 [1:09:13<14:10:35, 17.05s/it]2026-04-12 11:55:25,708 | INFO | train_grpo_train | metrics_logged mode=train step=306

                                                       
{'loss': 0.096, 'grad_norm': 15.95866584777832, 'learning_rate': 9.075757575757577e-06, 'num_tokens': 614926.0, 'completions/mean_length': 37.0, 'completions/min_length': 27.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.0, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.9003474116325378, 'rewards/meter/std': 0.13265357911586761, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5375000238418579, 'rewards/judge_quality/std': 0.18077215552330017, 'rewards/repeat_penalty/mean': 0.9292006492614746, 'rewards/repeat_penalty/std': 0.12348253279924393, 'rewards/near_duplicate_penalty/mean': 0.9654132127761841, 'rewards/near_duplicate_penalty/std': 0.034489717334508896, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9893162250518799, 'rewards/distinct_2/std': 0.03021823801100254, 'rewards/total_composite/mean': 0.4877563714981079, 'rewards/total_composite/std': 0.1906237006187439, 'reward': 0.4877563714981079, 'reward_std': 0.1906237006187439, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2088146060705185, 'sampling/sampling_logp_difference/max': 1.240828037261963, 'sampling/importance_sampling_ratio/min': 0.28914472460746765, 'sampling/importance_sampling_ratio/mean': 1.029973030090332, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7366817593574524, 'clip_ratio/low_mean': 0.09791644988581538, 'clip_ratio/low_min': 0.09791644988581538, 'clip_ratio/high_mean': 0.07956166565418243, 'clip_ratio/high_max': 0.07956166565418243, 'clip_ratio/region_mean': 0.17747811553999782, 'reward_total_mean': 0.4877563714981079, 'reward_meter_mean': 0.9003474116325378, 'reward_meter_std': 0.13265357911586761, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9292006492614746, 'reward_repeat_penalty_std': 0.12348253279924393, 'reward_near_duplicate_penalty_mean': 0.9654132127761841, 'reward_near_duplicate_penalty_std': 0.034489717334508896, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9893162250518799, 'reward_distinct_2_std': 0.03021823801100254, 'reward_judge_quality_mean': 0.5375000238418579, 'reward_judge_quality_std': 0.18077215552330017, 'reward_total_composite_mean': 0.4877563714981079, 'reward_total_composite_std': 0.1906237006187439, 'epoch': 0.01}

  9%|▉         | 306/3300 [1:09:13<14:10:35, 17.05s/it]INFO 04-12 11:55:25 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 307/3300 [1:09:22<12:06:56, 14.57s/it]2026-04-12 11:55:34,510 | INFO | train_grpo_train | metrics_logged mode=train step=307

                                                       
{'loss': 0.036, 'grad_norm': 19.920801162719727, 'learning_rate': 9.072727272727273e-06, 'num_tokens': 616250.0, 'completions/mean_length': 21.5, 'completions/min_length': 17.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.5, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.830586314201355, 'rewards/meter/std': 0.3387489318847656, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5425000190734863, 'rewards/judge_quality/std': 0.34087073802948, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4118979275226593, 'rewards/total_composite/std': 0.3443230390548706, 'reward': 0.4118979275226593, 'reward_std': 0.3443230390548706, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20045839250087738, 'sampling/sampling_logp_difference/max': 1.885471224784851, 'sampling/importance_sampling_ratio/min': 0.15175753831863403, 'sampling/importance_sampling_ratio/mean': 1.0101442337036133, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5795673578977585, 'clip_ratio/low_mean': 0.08071046741679311, 'clip_ratio/low_min': 0.08071046741679311, 'clip_ratio/high_mean': 0.06173687521368265, 'clip_ratio/high_max': 0.06173687521368265, 'clip_ratio/region_mean': 0.14244734263047576, 'reward_total_mean': 0.4118979275226593, 'reward_meter_mean': 0.830586314201355, 'reward_meter_std': 0.3387489318847656, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5425000190734863, 'reward_judge_quality_std': 0.34087073802948, 'reward_total_composite_mean': 0.4118979275226593, 'reward_total_composite_std': 0.3443230390548706, 'epoch': 0.01}

  9%|▉         | 307/3300 [1:09:22<12:06:56, 14.57s/it]INFO 04-12 11:55:34 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 308/3300 [1:09:37<12:12:13, 14.68s/it]2026-04-12 11:55:49,453 | INFO | train_grpo_train | metrics_logged mode=train step=308

                                                       
{'loss': -0.1, 'grad_norm': 3.589592933654785, 'learning_rate': 9.06969696969697e-06, 'num_tokens': 619110.0, 'completions/mean_length': 452.5, 'completions/min_length': 311.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.625, 'completions/mean_terminated_length': 353.3333435058594, 'completions/min_terminated_length': 311.0, 'completions/max_terminated_length': 421.0, 'rewards/meter/mean': 0.4821336269378662, 'rewards/meter/std': 0.270393431186676, 'rewards/count_adherence/mean': 0.6736111640930176, 'rewards/count_adherence/std': 0.24386203289031982, 'rewards/arabic_clean/mean': 0.25, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.8378704190254211, 'rewards/lexical_plausibility/std': 0.13672730326652527, 'rewards/judge_quality/mean': 0.11249999701976776, 'rewards/judge_quality/std': 0.051754921674728394, 'rewards/repeat_penalty/mean': 0.9791123867034912, 'rewards/repeat_penalty/std': 0.022714849561452866, 'rewards/near_duplicate_penalty/mean': 0.9937890768051147, 'rewards/near_duplicate_penalty/std': 0.004387399181723595, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9851706624031067, 'rewards/distinct_2/std': 0.019177502021193504, 'rewards/total_composite/mean': 0.019215591251850128, 'rewards/total_composite/std': 0.04168100655078888, 'reward': 0.019215591251850128, 'reward_std': 0.04168100282549858, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21137875318527222, 'sampling/sampling_logp_difference/max': 1.6428661346435547, 'sampling/importance_sampling_ratio/min': 0.19342486560344696, 'sampling/importance_sampling_ratio/mean': 1.0538774728775024, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9971690475940704, 'clip_ratio/low_mean': 0.028733489103615284, 'clip_ratio/low_min': 0.028733489103615284, 'clip_ratio/high_mean': 0.021704180166125298, 'clip_ratio/high_max': 0.021704180166125298, 'clip_ratio/region_mean': 0.05043766926974058, 'reward_total_mean': 0.019215591251850128, 'reward_meter_mean': 0.4821336269378662, 'reward_meter_std': 0.270393431186676, 'reward_count_adherence_mean': 0.6736111640930176, 'reward_count_adherence_std': 0.24386203289031982, 'reward_arabic_clean_mean': 0.25, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.8378704190254211, 'reward_lexical_plausibility_std': 0.13672730326652527, 'reward_repeat_penalty_mean': 0.9791123867034912, 'reward_repeat_penalty_std': 0.022714849561452866, 'reward_near_duplicate_penalty_mean': 0.9937890768051147, 'reward_near_duplicate_penalty_std': 0.004387399181723595, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9851706624031067, 'reward_distinct_2_std': 0.019177502021193504, 'reward_judge_quality_mean': 0.11249999701976776, 'reward_judge_quality_std': 0.051754921674728394, 'reward_total_composite_mean': 0.019215591251850128, 'reward_total_composite_std': 0.04168100655078888, 'epoch': 0.01}

  9%|▉         | 308/3300 [1:09:37<12:12:13, 14.68s/it]INFO 04-12 11:55:49 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 309/3300 [1:09:52<12:22:41, 14.90s/it]2026-04-12 11:56:04,851 | INFO | train_grpo_train | metrics_logged mode=train step=309

                                                       
{'loss': -0.0789, 'grad_norm': 4.695291519165039, 'learning_rate': 9.066666666666667e-06, 'num_tokens': 621724.0, 'completions/mean_length': 244.75, 'completions/min_length': 131.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 155.6666717529297, 'completions/min_terminated_length': 131.0, 'completions/max_terminated_length': 179.0, 'rewards/meter/mean': 0.4317730665206909, 'rewards/meter/std': 0.4164171814918518, 'rewards/count_adherence/mean': 0.890625, 'rewards/count_adherence/std': 0.14074896275997162, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9328837394714355, 'rewards/lexical_plausibility/std': 0.12524037063121796, 'rewards/judge_quality/mean': 0.29999998211860657, 'rewards/judge_quality/std': 0.16903084516525269, 'rewards/repeat_penalty/mean': 0.9769273996353149, 'rewards/repeat_penalty/std': 0.03193797543644905, 'rewards/near_duplicate_penalty/mean': 0.9913658499717712, 'rewards/near_duplicate_penalty/std': 0.005720204208046198, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9853183627128601, 'rewards/distinct_2/std': 0.027391478419303894, 'rewards/total_composite/mean': 0.14949725568294525, 'rewards/total_composite/std': 0.18220500648021698, 'reward': 0.14949725568294525, 'reward_std': 0.18220500648021698, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2315247803926468, 'sampling/sampling_logp_difference/max': 1.192319631576538, 'sampling/importance_sampling_ratio/min': 0.34306013584136963, 'sampling/importance_sampling_ratio/mean': 1.0498822927474976, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.3267560601234436, 'clip_ratio/low_mean': 0.0709469635039568, 'clip_ratio/low_min': 0.0709469635039568, 'clip_ratio/high_mean': 0.07881752587854862, 'clip_ratio/high_max': 0.07881752587854862, 'clip_ratio/region_mean': 0.14976448938250542, 'reward_total_mean': 0.14949725568294525, 'reward_meter_mean': 0.4317730665206909, 'reward_meter_std': 0.4164171814918518, 'reward_count_adherence_mean': 0.890625, 'reward_count_adherence_std': 0.14074896275997162, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9328837394714355, 'reward_lexical_plausibility_std': 0.12524037063121796, 'reward_repeat_penalty_mean': 0.9769273996353149, 'reward_repeat_penalty_std': 0.03193797543644905, 'reward_near_duplicate_penalty_mean': 0.9913658499717712, 'reward_near_duplicate_penalty_std': 0.005720204208046198, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9853183627128601, 'reward_distinct_2_std': 0.027391478419303894, 'reward_judge_quality_mean': 0.29999998211860657, 'reward_judge_quality_std': 0.16903084516525269, 'reward_total_composite_mean': 0.14949725568294525, 'reward_total_composite_std': 0.18220500648021698, 'epoch': 0.01}

  9%|▉         | 309/3300 [1:09:52<12:22:41, 14.90s/it]INFO 04-12 11:56:05 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 310/3300 [1:10:02<11:05:10, 13.35s/it]2026-04-12 11:56:14,583 | INFO | train_grpo_train | metrics_logged mode=train step=310

                                                       
{'loss': -0.0056, 'grad_norm': 14.204791069030762, 'learning_rate': 9.063636363636365e-06, 'num_tokens': 623684.0, 'completions/mean_length': 57.0, 'completions/min_length': 50.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.0, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.8325011730194092, 'rewards/meter/std': 0.3362966775894165, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9944852590560913, 'rewards/lexical_plausibility/std': 0.015597938559949398, 'rewards/judge_quality/mean': 0.5049999952316284, 'rewards/judge_quality/std': 0.2745385766029358, 'rewards/repeat_penalty/mean': 0.9572106599807739, 'rewards/repeat_penalty/std': 0.038330599665641785, 'rewards/near_duplicate_penalty/mean': 0.9769012928009033, 'rewards/near_duplicate_penalty/std': 0.016843775287270546, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9796324372291565, 'rewards/distinct_2/std': 0.02811598777770996, 'rewards/total_composite/mean': 0.46715015172958374, 'rewards/total_composite/std': 0.3051133453845978, 'reward': 0.46715015172958374, 'reward_std': 0.3051133453845978, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2098153829574585, 'sampling/sampling_logp_difference/max': 1.2948904037475586, 'sampling/importance_sampling_ratio/min': 0.2739278972148895, 'sampling/importance_sampling_ratio/mean': 1.0542532205581665, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9368436485528946, 'clip_ratio/low_mean': 0.13437646813690662, 'clip_ratio/low_min': 0.13437646813690662, 'clip_ratio/high_mean': 0.034722222946584225, 'clip_ratio/high_max': 0.034722222946584225, 'clip_ratio/region_mean': 0.16909869108349085, 'reward_total_mean': 0.46715015172958374, 'reward_meter_mean': 0.8325011730194092, 'reward_meter_std': 0.3362966775894165, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9944852590560913, 'reward_lexical_plausibility_std': 0.015597938559949398, 'reward_repeat_penalty_mean': 0.9572106599807739, 'reward_repeat_penalty_std': 0.038330599665641785, 'reward_near_duplicate_penalty_mean': 0.9769012928009033, 'reward_near_duplicate_penalty_std': 0.016843775287270546, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9796324372291565, 'reward_distinct_2_std': 0.02811598777770996, 'reward_judge_quality_mean': 0.5049999952316284, 'reward_judge_quality_std': 0.2745385766029358, 'reward_total_composite_mean': 0.46715015172958374, 'reward_total_composite_std': 0.3051133453845978, 'epoch': 0.01}

  9%|▉         | 310/3300 [1:10:02<11:05:10, 13.35s/it]INFO 04-12 11:56:14 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 311/3300 [1:10:12<10:18:54, 12.42s/it]2026-04-12 11:56:24,850 | INFO | train_grpo_train | metrics_logged mode=train step=311

                                                       
{'loss': 0.0685, 'grad_norm': 13.525808334350586, 'learning_rate': 9.06060606060606e-06, 'num_tokens': 625707.0, 'completions/mean_length': 79.875, 'completions/min_length': 61.0, 'completions/max_length': 95.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.875, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 95.0, 'rewards/meter/mean': 0.39014071226119995, 'rewards/meter/std': 0.2583003640174866, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9834882020950317, 'rewards/lexical_plausibility/std': 0.04670235514640808, 'rewards/judge_quality/mean': 0.36249998211860657, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.974642276763916, 'rewards/repeat_penalty/std': 0.029900040477514267, 'rewards/near_duplicate_penalty/mean': 0.9878274202346802, 'rewards/near_duplicate_penalty/std': 0.007417390588670969, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9865831732749939, 'rewards/distinct_2/std': 0.0266198068857193, 'rewards/total_composite/mean': 0.12803345918655396, 'rewards/total_composite/std': 0.11432408541440964, 'reward': 0.12803345918655396, 'reward_std': 0.11432407051324844, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20194865763187408, 'sampling/sampling_logp_difference/max': 2.013035774230957, 'sampling/importance_sampling_ratio/min': 0.13358253240585327, 'sampling/importance_sampling_ratio/mean': 1.0232917070388794, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5410507768392563, 'clip_ratio/low_mean': 0.08136198483407497, 'clip_ratio/low_min': 0.08136198483407497, 'clip_ratio/high_mean': 0.09056342486292124, 'clip_ratio/high_max': 0.09056342486292124, 'clip_ratio/region_mean': 0.1719254096969962, 'reward_total_mean': 0.12803345918655396, 'reward_meter_mean': 0.39014071226119995, 'reward_meter_std': 0.2583003640174866, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.0707106739282608, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9834882020950317, 'reward_lexical_plausibility_std': 0.04670235514640808, 'reward_repeat_penalty_mean': 0.974642276763916, 'reward_repeat_penalty_std': 0.029900040477514267, 'reward_near_duplicate_penalty_mean': 0.9878274202346802, 'reward_near_duplicate_penalty_std': 0.007417390588670969, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9865831732749939, 'reward_distinct_2_std': 0.0266198068857193, 'reward_judge_quality_mean': 0.36249998211860657, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.12803345918655396, 'reward_total_composite_std': 0.11432408541440964, 'epoch': 0.01}

  9%|▉         | 311/3300 [1:10:12<10:18:54, 12.42s/it]INFO 04-12 11:56:25 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 312/3300 [1:10:27<10:45:58, 12.97s/it]2026-04-12 11:56:39,098 | INFO | train_grpo_train | metrics_logged mode=train step=312

                                                       
{'loss': -0.1229, 'grad_norm': 5.748407363891602, 'learning_rate': 9.057575757575759e-06, 'num_tokens': 627468.0, 'completions/mean_length': 111.125, 'completions/min_length': 50.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 53.85714340209961, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.827028214931488, 'rewards/meter/std': 0.3209535777568817, '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/lexical_plausibility/mean': 0.9249270558357239, 'rewards/lexical_plausibility/std': 0.21233832836151123, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.1922609955072403, 'rewards/repeat_penalty/mean': 0.8958534002304077, 'rewards/repeat_penalty/std': 0.15997065603733063, 'rewards/near_duplicate_penalty/mean': 0.9642992615699768, 'rewards/near_duplicate_penalty/std': 0.031497787684202194, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9450549483299255, 'rewards/distinct_2/std': 0.10173846781253815, 'rewards/total_composite/mean': 0.3584931492805481, 'rewards/total_composite/std': 0.2006467878818512, 'reward': 0.3584931492805481, 'reward_std': 0.20064677298069, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23162570595741272, 'sampling/sampling_logp_difference/max': 1.6000958681106567, 'sampling/importance_sampling_ratio/min': 0.201877161860466, 'sampling/importance_sampling_ratio/mean': 1.0524348020553589, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8471671640872955, 'clip_ratio/low_mean': 0.1423076931387186, 'clip_ratio/low_min': 0.1423076931387186, 'clip_ratio/high_mean': 0.04150943458080292, 'clip_ratio/high_max': 0.04150943458080292, 'clip_ratio/region_mean': 0.18381712771952152, 'reward_total_mean': 0.3584931492805481, 'reward_meter_mean': 0.827028214931488, 'reward_meter_std': 0.3209535777568817, '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_lexical_plausibility_mean': 0.9249270558357239, 'reward_lexical_plausibility_std': 0.21233832836151123, 'reward_repeat_penalty_mean': 0.8958534002304077, 'reward_repeat_penalty_std': 0.15997065603733063, 'reward_near_duplicate_penalty_mean': 0.9642992615699768, 'reward_near_duplicate_penalty_std': 0.031497787684202194, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9450549483299255, 'reward_distinct_2_std': 0.10173846781253815, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.1922609955072403, 'reward_total_composite_mean': 0.3584931492805481, 'reward_total_composite_std': 0.2006467878818512, 'epoch': 0.01}

  9%|▉         | 312/3300 [1:10:27<10:45:58, 12.97s/it]INFO 04-12 11:56:39 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 313/3300 [1:10:36<10:00:21, 12.06s/it]2026-04-12 11:56:49,031 | INFO | train_grpo_train | metrics_logged mode=train step=313

                                                       
{'loss': 0.0259, 'grad_norm': 14.188372611999512, 'learning_rate': 9.054545454545455e-06, 'num_tokens': 629439.0, 'completions/mean_length': 79.375, 'completions/min_length': 68.0, 'completions/max_length': 84.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.375, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.7856276035308838, 'rewards/meter/std': 0.3154408931732178, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.42499998211860657, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.9707048535346985, 'rewards/repeat_penalty/std': 0.031784094870090485, 'rewards/near_duplicate_penalty/mean': 0.9863473773002625, 'rewards/near_duplicate_penalty/std': 0.01276310719549656, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9839661121368408, 'rewards/distinct_2/std': 0.022137979045510292, 'rewards/total_composite/mean': 0.3288583755493164, 'rewards/total_composite/std': 0.130870059132576, 'reward': 0.3288583755493164, 'reward_std': 0.130870059132576, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17699967324733734, 'sampling/sampling_logp_difference/max': 1.2040987014770508, 'sampling/importance_sampling_ratio/min': 0.29996222257614136, 'sampling/importance_sampling_ratio/mean': 1.0302438735961914, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6164602041244507, 'clip_ratio/low_mean': 0.04212055169045925, 'clip_ratio/low_min': 0.04212055169045925, 'clip_ratio/high_mean': 0.12164961267262697, 'clip_ratio/high_max': 0.12164961267262697, 'clip_ratio/region_mean': 0.16377016436308622, 'reward_total_mean': 0.3288583755493164, 'reward_meter_mean': 0.7856276035308838, 'reward_meter_std': 0.3154408931732178, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9707048535346985, 'reward_repeat_penalty_std': 0.031784094870090485, 'reward_near_duplicate_penalty_mean': 0.9863473773002625, 'reward_near_duplicate_penalty_std': 0.01276310719549656, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9839661121368408, 'reward_distinct_2_std': 0.022137979045510292, 'reward_judge_quality_mean': 0.42499998211860657, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.3288583755493164, 'reward_total_composite_std': 0.130870059132576, 'epoch': 0.01}

  9%|▉         | 313/3300 [1:10:37<10:00:21, 12.06s/it]INFO 04-12 11:56:49 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 314/3300 [1:10:51<10:36:47, 12.80s/it]2026-04-12 11:57:03,560 | INFO | train_grpo_train | metrics_logged mode=train step=314

                                                       
{'loss': -0.1353, 'grad_norm': 3.3886120319366455, 'learning_rate': 9.051515151515152e-06, 'num_tokens': 632135.0, 'completions/mean_length': 306.0, 'completions/min_length': 163.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 182.40000915527344, 'completions/min_terminated_length': 163.0, 'completions/max_terminated_length': 221.0, 'rewards/meter/mean': 0.6670107841491699, 'rewards/meter/std': 0.30283650755882263, 'rewards/count_adherence/mean': 0.7777777910232544, 'rewards/count_adherence/std': 0.2519763112068176, 'rewards/arabic_clean/mean': 0.375, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.8064969778060913, 'rewards/lexical_plausibility/std': 0.23627878725528717, 'rewards/judge_quality/mean': 0.21250000596046448, 'rewards/judge_quality/std': 0.1597989797592163, 'rewards/repeat_penalty/mean': 0.9876633882522583, 'rewards/repeat_penalty/std': 0.014699472114443779, 'rewards/near_duplicate_penalty/mean': 0.9937053322792053, 'rewards/near_duplicate_penalty/std': 0.005136602092534304, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.99388587474823, 'rewards/distinct_2/std': 0.010835127905011177, 'rewards/total_composite/mean': 0.0810789167881012, 'rewards/total_composite/std': 0.13296110928058624, 'reward': 0.0810789167881012, 'reward_std': 0.13296109437942505, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20112429559230804, 'sampling/sampling_logp_difference/max': 1.3474502563476562, 'sampling/importance_sampling_ratio/min': 0.25990211963653564, 'sampling/importance_sampling_ratio/mean': 1.0480228662490845, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0854735374450684, 'clip_ratio/low_mean': 0.03859323728829622, 'clip_ratio/low_min': 0.03859323728829622, 'clip_ratio/high_mean': 0.06395599246025085, 'clip_ratio/high_max': 0.06395599246025085, 'clip_ratio/region_mean': 0.10254922974854708, 'reward_total_mean': 0.0810789167881012, 'reward_meter_mean': 0.6670107841491699, 'reward_meter_std': 0.30283650755882263, 'reward_count_adherence_mean': 0.7777777910232544, 'reward_count_adherence_std': 0.2519763112068176, 'reward_arabic_clean_mean': 0.375, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.8064969778060913, 'reward_lexical_plausibility_std': 0.23627878725528717, 'reward_repeat_penalty_mean': 0.9876633882522583, 'reward_repeat_penalty_std': 0.014699472114443779, 'reward_near_duplicate_penalty_mean': 0.9937053322792053, 'reward_near_duplicate_penalty_std': 0.005136602092534304, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.99388587474823, 'reward_distinct_2_std': 0.010835127905011177, 'reward_judge_quality_mean': 0.21250000596046448, 'reward_judge_quality_std': 0.1597989797592163, 'reward_total_composite_mean': 0.0810789167881012, 'reward_total_composite_std': 0.13296110928058624, 'epoch': 0.01}

 10%|▉         | 314/3300 [1:10:51<10:36:47, 12.80s/it]INFO 04-12 11:57:03 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 315/3300 [1:11:01<9:53:50, 11.94s/it] 2026-04-12 11:57:13,478 | INFO | train_grpo_train | metrics_logged mode=train step=315

                                                      
{'loss': 0.0604, 'grad_norm': 14.44451904296875, 'learning_rate': 9.04848484848485e-06, 'num_tokens': 633894.0, 'completions/mean_length': 52.875, 'completions/min_length': 44.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.875, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.8267420530319214, 'rewards/meter/std': 0.3160305619239807, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5674999952316284, 'rewards/judge_quality/std': 0.21756774187088013, 'rewards/repeat_penalty/mean': 0.90375816822052, 'rewards/repeat_penalty/std': 0.08299682289361954, 'rewards/near_duplicate_penalty/mean': 0.9495149254798889, 'rewards/near_duplicate_penalty/std': 0.029489515349268913, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9513822197914124, 'rewards/distinct_2/std': 0.07583235204219818, 'rewards/total_composite/mean': 0.4872320294380188, 'rewards/total_composite/std': 0.2897428274154663, 'reward': 0.4872320294380188, 'reward_std': 0.2897427976131439, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16570477187633514, 'sampling/sampling_logp_difference/max': 2.175581932067871, 'sampling/importance_sampling_ratio/min': 0.11354205757379532, 'sampling/importance_sampling_ratio/mean': 1.0141010284423828, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1112485229969025, 'clip_ratio/low_mean': 0.11848831083625555, 'clip_ratio/low_min': 0.11848831083625555, 'clip_ratio/high_mean': 0.04618818685412407, 'clip_ratio/high_max': 0.04618818685412407, 'clip_ratio/region_mean': 0.16467649769037962, 'reward_total_mean': 0.4872320294380188, 'reward_meter_mean': 0.8267420530319214, 'reward_meter_std': 0.3160305619239807, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.90375816822052, 'reward_repeat_penalty_std': 0.08299682289361954, 'reward_near_duplicate_penalty_mean': 0.9495149254798889, 'reward_near_duplicate_penalty_std': 0.029489515349268913, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9513822197914124, 'reward_distinct_2_std': 0.07583235204219818, 'reward_judge_quality_mean': 0.5674999952316284, 'reward_judge_quality_std': 0.21756774187088013, 'reward_total_composite_mean': 0.4872320294380188, 'reward_total_composite_std': 0.2897428274154663, 'epoch': 0.01}

 10%|▉         | 315/3300 [1:11:01<9:53:50, 11.94s/it]INFO 04-12 11:57:13 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 316/3300 [1:11:15<10:27:53, 12.63s/it]2026-04-12 11:57:27,720 | INFO | train_grpo_train | metrics_logged mode=train step=316

                                                       
{'loss': -0.1272, 'grad_norm': 3.496385335922241, 'learning_rate': 9.045454545454546e-06, 'num_tokens': 636364.0, 'completions/mean_length': 240.75, 'completions/min_length': 120.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 150.33334350585938, 'completions/min_terminated_length': 120.0, 'completions/max_terminated_length': 221.0, 'rewards/meter/mean': 0.6570875644683838, 'rewards/meter/std': 0.3701251149177551, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.34503278136253357, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.8570929169654846, 'rewards/lexical_plausibility/std': 0.2655525207519531, 'rewards/judge_quality/mean': 0.29999998211860657, 'rewards/judge_quality/std': 0.18516401946544647, 'rewards/repeat_penalty/mean': 0.9834110736846924, 'rewards/repeat_penalty/std': 0.021247945725917816, 'rewards/near_duplicate_penalty/mean': 0.9942145347595215, 'rewards/near_duplicate_penalty/std': 0.004308826755732298, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9891218543052673, 'rewards/distinct_2/std': 0.020289666950702667, 'rewards/total_composite/mean': 0.21258118748664856, 'rewards/total_composite/std': 0.17661871016025543, 'reward': 0.21258118748664856, 'reward_std': 0.17661871016025543, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20938053727149963, 'sampling/sampling_logp_difference/max': 1.2175943851470947, 'sampling/importance_sampling_ratio/min': 0.29594123363494873, 'sampling/importance_sampling_ratio/mean': 1.0499805212020874, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6157505214214325, 'clip_ratio/low_mean': 0.04339649248868227, 'clip_ratio/low_min': 0.04339649248868227, 'clip_ratio/high_mean': 0.10528704524040222, 'clip_ratio/high_max': 0.10528704524040222, 'clip_ratio/region_mean': 0.1486835377290845, 'reward_total_mean': 0.21258118748664856, 'reward_meter_mean': 0.6570875644683838, 'reward_meter_std': 0.3701251149177551, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.34503278136253357, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.8570929169654846, 'reward_lexical_plausibility_std': 0.2655525207519531, 'reward_repeat_penalty_mean': 0.9834110736846924, 'reward_repeat_penalty_std': 0.021247945725917816, 'reward_near_duplicate_penalty_mean': 0.9942145347595215, 'reward_near_duplicate_penalty_std': 0.004308826755732298, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9891218543052673, 'reward_distinct_2_std': 0.020289666950702667, 'reward_judge_quality_mean': 0.29999998211860657, 'reward_judge_quality_std': 0.18516401946544647, 'reward_total_composite_mean': 0.21258118748664856, 'reward_total_composite_std': 0.17661871016025543, 'epoch': 0.01}

 10%|▉         | 316/3300 [1:11:15<10:27:53, 12.63s/it]INFO 04-12 11:57:28 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 317/3300 [1:11:24<9:32:12, 11.51s/it] 2026-04-12 11:57:36,615 | INFO | train_grpo_train | metrics_logged mode=train step=317

                                                      
{'loss': -0.1125, 'grad_norm': 17.743989944458008, 'learning_rate': 9.042424242424244e-06, 'num_tokens': 637907.0, 'completions/mean_length': 20.875, 'completions/min_length': 17.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.875, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.9689522981643677, 'rewards/meter/std': 0.005783551372587681, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5425000190734863, 'rewards/judge_quality/std': 0.32805708050727844, 'rewards/repeat_penalty/mean': 0.7307312488555908, 'rewards/repeat_penalty/std': 0.02059919387102127, 'rewards/near_duplicate_penalty/mean': 0.9743083119392395, 'rewards/near_duplicate_penalty/std': 0.027465570718050003, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5245275497436523, 'rewards/total_composite/std': 0.31540247797966003, 'reward': 0.5245275497436523, 'reward_std': 0.31540247797966003, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1126757487654686, 'sampling/sampling_logp_difference/max': 1.3009719848632812, 'sampling/importance_sampling_ratio/min': 0.3247632384300232, 'sampling/importance_sampling_ratio/mean': 1.033617615699768, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5779096521437168, 'clip_ratio/low_mean': 0.08270369609817863, 'clip_ratio/low_min': 0.08270369609817863, 'clip_ratio/high_mean': 0.04456018563359976, 'clip_ratio/high_max': 0.04456018563359976, 'clip_ratio/region_mean': 0.12726388173177838, 'reward_total_mean': 0.5245275497436523, 'reward_meter_mean': 0.9689522981643677, 'reward_meter_std': 0.005783551372587681, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7307312488555908, 'reward_repeat_penalty_std': 0.02059919387102127, 'reward_near_duplicate_penalty_mean': 0.9743083119392395, 'reward_near_duplicate_penalty_std': 0.027465570718050003, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5425000190734863, 'reward_judge_quality_std': 0.32805708050727844, 'reward_total_composite_mean': 0.5245275497436523, 'reward_total_composite_std': 0.31540247797966003, 'epoch': 0.01}

 10%|▉         | 317/3300 [1:11:24<9:32:12, 11.51s/it]INFO 04-12 11:57:36 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 318/3300 [1:11:33<8:56:50, 10.80s/it]2026-04-12 11:57:45,768 | INFO | train_grpo_train | metrics_logged mode=train step=318

                                                      
{'loss': 0.0252, 'grad_norm': 17.594728469848633, 'learning_rate': 9.03939393939394e-06, 'num_tokens': 639659.0, 'completions/mean_length': 44.0, 'completions/min_length': 40.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.0, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.5897473096847534, 'rewards/meter/std': 0.3423965573310852, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5887500047683716, 'rewards/judge_quality/std': 0.27740830183029175, 'rewards/repeat_penalty/mean': 0.9881412386894226, 'rewards/repeat_penalty/std': 0.01636464148759842, 'rewards/near_duplicate_penalty/mean': 0.9881412386894226, 'rewards/near_duplicate_penalty/std': 0.01636464148759842, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.37728744745254517, 'rewards/total_composite/std': 0.30481475591659546, 'reward': 0.37728744745254517, 'reward_std': 0.30481472611427307, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1902374029159546, 'sampling/sampling_logp_difference/max': 1.3638861179351807, 'sampling/importance_sampling_ratio/min': 0.25566530227661133, 'sampling/importance_sampling_ratio/mean': 1.0728695392608643, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7821234613656998, 'clip_ratio/low_mean': 0.10526788979768753, 'clip_ratio/low_min': 0.10526788979768753, 'clip_ratio/high_mean': 0.08607660513371229, 'clip_ratio/high_max': 0.08607660513371229, 'clip_ratio/region_mean': 0.19134449493139982, 'reward_total_mean': 0.37728744745254517, 'reward_meter_mean': 0.5897473096847534, 'reward_meter_std': 0.3423965573310852, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9881412386894226, 'reward_repeat_penalty_std': 0.01636464148759842, 'reward_near_duplicate_penalty_mean': 0.9881412386894226, 'reward_near_duplicate_penalty_std': 0.01636464148759842, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5887500047683716, 'reward_judge_quality_std': 0.27740830183029175, 'reward_total_composite_mean': 0.37728744745254517, 'reward_total_composite_std': 0.30481475591659546, 'epoch': 0.01}

 10%|▉         | 318/3300 [1:11:33<8:56:50, 10.80s/it]INFO 04-12 11:57:46 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 319/3300 [1:11:49<10:08:37, 12.25s/it]2026-04-12 11:58:01,395 | INFO | train_grpo_train | metrics_logged mode=train step=319

                                                       
{'loss': -0.081, 'grad_norm': 4.2549848556518555, 'learning_rate': 9.036363636363638e-06, 'num_tokens': 641286.0, 'completions/mean_length': 97.375, 'completions/min_length': 35.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 38.142860412597656, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.9505184888839722, 'rewards/meter/std': 0.08344382047653198, '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/lexical_plausibility/mean': 0.9379931092262268, 'rewards/lexical_plausibility/std': 0.17538200318813324, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.1642080545425415, 'rewards/repeat_penalty/mean': 0.9991400241851807, 'rewards/repeat_penalty/std': 0.001592559041455388, 'rewards/near_duplicate_penalty/mean': 0.9991400241851807, 'rewards/near_duplicate_penalty/std': 0.001592559041455388, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.33270570635795593, 'rewards/total_composite/std': 0.20538638532161713, 'reward': 0.33270570635795593, 'reward_std': 0.20538638532161713, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21220289170742035, 'sampling/sampling_logp_difference/max': 1.1251039505004883, 'sampling/importance_sampling_ratio/min': 0.32461872696876526, 'sampling/importance_sampling_ratio/mean': 1.0525115728378296, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9547057002782822, 'clip_ratio/low_mean': 0.03928571566939354, 'clip_ratio/low_min': 0.03928571566939354, 'clip_ratio/high_mean': 0.15886845905333757, 'clip_ratio/high_max': 0.15886845905333757, 'clip_ratio/region_mean': 0.1981541747227311, 'reward_total_mean': 0.33270570635795593, 'reward_meter_mean': 0.9505184888839722, 'reward_meter_std': 0.08344382047653198, '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_lexical_plausibility_mean': 0.9379931092262268, 'reward_lexical_plausibility_std': 0.17538200318813324, 'reward_repeat_penalty_mean': 0.9991400241851807, 'reward_repeat_penalty_std': 0.001592559041455388, 'reward_near_duplicate_penalty_mean': 0.9991400241851807, 'reward_near_duplicate_penalty_std': 0.001592559041455388, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.1642080545425415, 'reward_total_composite_mean': 0.33270570635795593, 'reward_total_composite_std': 0.20538638532161713, 'epoch': 0.01}

 10%|▉         | 319/3300 [1:11:49<10:08:37, 12.25s/it]INFO 04-12 11:58:01 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 320/3300 [1:12:02<10:20:28, 12.49s/it]2026-04-12 11:58:14,459 | INFO | train_grpo_train | metrics_logged mode=train step=320

                                                       
{'loss': -0.0336, 'grad_norm': 3.643080949783325, 'learning_rate': 9.033333333333334e-06, 'num_tokens': 642738.0, 'completions/mean_length': 155.5, 'completions/min_length': 29.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 36.66666793823242, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.35714009404182434, 'rewards/meter/std': 0.4069850742816925, '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/lexical_plausibility/mean': 0.8637242317199707, 'rewards/lexical_plausibility/std': 0.23499347269535065, 'rewards/judge_quality/mean': 0.44875001907348633, 'rewards/judge_quality/std': 0.3410671353340149, 'rewards/repeat_penalty/mean': 0.98957359790802, 'rewards/repeat_penalty/std': 0.019379448145627975, 'rewards/near_duplicate_penalty/mean': 0.98957359790802, 'rewards/near_duplicate_penalty/std': 0.019379448145627975, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.19615322351455688, 'rewards/total_composite/std': 0.2576081156730652, 'reward': 0.19615322351455688, 'reward_std': 0.2576081156730652, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20727041363716125, 'sampling/sampling_logp_difference/max': 2.490852117538452, 'sampling/importance_sampling_ratio/min': 0.08283934742212296, 'sampling/importance_sampling_ratio/mean': 1.0229649543762207, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5510234534740448, 'clip_ratio/low_mean': 0.06580646149814129, 'clip_ratio/low_min': 0.06580646149814129, 'clip_ratio/high_mean': 0.06170794926583767, 'clip_ratio/high_max': 0.06170794926583767, 'clip_ratio/region_mean': 0.12751441076397896, 'reward_total_mean': 0.19615322351455688, 'reward_meter_mean': 0.35714009404182434, 'reward_meter_std': 0.4069850742816925, '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_lexical_plausibility_mean': 0.8637242317199707, 'reward_lexical_plausibility_std': 0.23499347269535065, 'reward_repeat_penalty_mean': 0.98957359790802, 'reward_repeat_penalty_std': 0.019379448145627975, 'reward_near_duplicate_penalty_mean': 0.98957359790802, 'reward_near_duplicate_penalty_std': 0.019379448145627975, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.44875001907348633, 'reward_judge_quality_std': 0.3410671353340149, 'reward_total_composite_mean': 0.19615322351455688, 'reward_total_composite_std': 0.2576081156730652, 'epoch': 0.01}

 10%|▉         | 320/3300 [1:12:02<10:20:28, 12.49s/it]INFO 04-12 11:58:14 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 321/3300 [1:12:10<9:14:48, 11.17s/it] 2026-04-12 11:58:22,554 | INFO | train_grpo_train | metrics_logged mode=train step=321

                                                      
{'loss': 0.0769, 'grad_norm': 22.280393600463867, 'learning_rate': 9.030303030303031e-06, 'num_tokens': 644240.0, 'completions/mean_length': 34.75, 'completions/min_length': 30.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.75, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.6890154480934143, 'rewards/meter/std': 0.4056783616542816, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8612500429153442, 'rewards/judge_quality/std': 0.16617010533809662, 'rewards/repeat_penalty/mean': 0.9721575379371643, 'rewards/repeat_penalty/std': 0.044762156903743744, 'rewards/near_duplicate_penalty/mean': 0.9817650318145752, 'rewards/near_duplicate_penalty/std': 0.023463450372219086, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9898785352706909, 'rewards/distinct_2/std': 0.028627805411815643, 'rewards/total_composite/mean': 0.5973183512687683, 'rewards/total_composite/std': 0.4225291311740875, 'reward': 0.5973183512687683, 'reward_std': 0.42252910137176514, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.185963436961174, 'sampling/sampling_logp_difference/max': 2.1489572525024414, 'sampling/importance_sampling_ratio/min': 0.11660567671060562, 'sampling/importance_sampling_ratio/mean': 1.0001929998397827, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2793263271450996, 'clip_ratio/low_mean': 0.07067149505019188, 'clip_ratio/low_min': 0.07067149505019188, 'clip_ratio/high_mean': 0.15500398725271225, 'clip_ratio/high_max': 0.15500398725271225, 'clip_ratio/region_mean': 0.22567548230290413, 'reward_total_mean': 0.5973183512687683, 'reward_meter_mean': 0.6890154480934143, 'reward_meter_std': 0.4056783616542816, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9721575379371643, 'reward_repeat_penalty_std': 0.044762156903743744, 'reward_near_duplicate_penalty_mean': 0.9817650318145752, 'reward_near_duplicate_penalty_std': 0.023463450372219086, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9898785352706909, 'reward_distinct_2_std': 0.028627805411815643, 'reward_judge_quality_mean': 0.8612500429153442, 'reward_judge_quality_std': 0.16617010533809662, 'reward_total_composite_mean': 0.5973183512687683, 'reward_total_composite_std': 0.4225291311740875, 'epoch': 0.01}

 10%|▉         | 321/3300 [1:12:10<9:14:48, 11.17s/it]INFO 04-12 11:58:22 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 322/3300 [1:12:21<9:19:02, 11.26s/it]2026-04-12 11:58:34,024 | INFO | train_grpo_train | metrics_logged mode=train step=322

                                                      
{'loss': 0.0675, 'grad_norm': 8.970571517944336, 'learning_rate': 9.027272727272728e-06, 'num_tokens': 646945.0, 'completions/mean_length': 147.125, 'completions/min_length': 120.0, 'completions/max_length': 162.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 147.125, 'completions/min_terminated_length': 120.0, 'completions/max_terminated_length': 162.0, 'rewards/meter/mean': 0.8744029402732849, 'rewards/meter/std': 0.21563386917114258, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0578637570142746, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.7203657627105713, 'rewards/repeat_penalty/std': 0.231900155544281, 'rewards/near_duplicate_penalty/mean': 0.9471039772033691, 'rewards/near_duplicate_penalty/std': 0.045894794166088104, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.7840245962142944, 'rewards/distinct_2/std': 0.19440589845180511, 'rewards/total_composite/mean': 0.31651216745376587, 'rewards/total_composite/std': 0.08478568494319916, 'reward': 0.31651216745376587, 'reward_std': 0.08478567749261856, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15659639239311218, 'sampling/sampling_logp_difference/max': 2.8806912899017334, 'sampling/importance_sampling_ratio/min': 0.056095968931913376, 'sampling/importance_sampling_ratio/mean': 1.0231878757476807, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2355171293020248, 'clip_ratio/low_mean': 0.043371642008423805, 'clip_ratio/low_min': 0.043371642008423805, 'clip_ratio/high_mean': 0.10302290134131908, 'clip_ratio/high_max': 0.10302290134131908, 'clip_ratio/region_mean': 0.1463945433497429, 'reward_total_mean': 0.31651216745376587, 'reward_meter_mean': 0.8744029402732849, 'reward_meter_std': 0.21563386917114258, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0578637570142746, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7203657627105713, 'reward_repeat_penalty_std': 0.231900155544281, 'reward_near_duplicate_penalty_mean': 0.9471039772033691, 'reward_near_duplicate_penalty_std': 0.045894794166088104, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.7840245962142944, 'reward_distinct_2_std': 0.19440589845180511, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.31651216745376587, 'reward_total_composite_std': 0.08478568494319916, 'epoch': 0.01}

 10%|▉         | 322/3300 [1:12:21<9:19:02, 11.26s/it]INFO 04-12 11:58:34 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 323/3300 [1:12:31<8:46:38, 10.61s/it]2026-04-12 11:58:43,133 | INFO | train_grpo_train | metrics_logged mode=train step=323

                                                      
{'loss': 0.0283, 'grad_norm': 14.265707969665527, 'learning_rate': 9.024242424242426e-06, 'num_tokens': 648538.0, 'completions/mean_length': 46.125, 'completions/min_length': 34.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.125, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9333639144897461, 'rewards/meter/std': 0.08309081941843033, '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/lexical_plausibility/mean': 0.9833590984344482, 'rewards/lexical_plausibility/std': 0.04706747457385063, 'rewards/judge_quality/mean': 0.7312500476837158, 'rewards/judge_quality/std': 0.3589046001434326, 'rewards/repeat_penalty/mean': 0.9403373003005981, 'rewards/repeat_penalty/std': 0.11561089009046555, 'rewards/near_duplicate_penalty/mean': 0.9874398112297058, 'rewards/near_duplicate_penalty/std': 0.013963574543595314, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9801586866378784, 'rewards/distinct_2/std': 0.036883533000946045, 'rewards/total_composite/mean': 0.6683478355407715, 'rewards/total_composite/std': 0.3816838264465332, 'reward': 0.6683478355407715, 'reward_std': 0.3816838264465332, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16488775610923767, 'sampling/sampling_logp_difference/max': 1.1753854751586914, 'sampling/importance_sampling_ratio/min': 0.30869996547698975, 'sampling/importance_sampling_ratio/mean': 1.0082929134368896, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1936316415667534, 'clip_ratio/low_mean': 0.04910714365541935, 'clip_ratio/low_min': 0.04910714365541935, 'clip_ratio/high_mean': 0.13271064218133688, 'clip_ratio/high_max': 0.13271064218133688, 'clip_ratio/region_mean': 0.18181778583675623, 'reward_total_mean': 0.6683478355407715, 'reward_meter_mean': 0.9333639144897461, 'reward_meter_std': 0.08309081941843033, '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_lexical_plausibility_mean': 0.9833590984344482, 'reward_lexical_plausibility_std': 0.04706747457385063, 'reward_repeat_penalty_mean': 0.9403373003005981, 'reward_repeat_penalty_std': 0.11561089009046555, 'reward_near_duplicate_penalty_mean': 0.9874398112297058, 'reward_near_duplicate_penalty_std': 0.013963574543595314, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9801586866378784, 'reward_distinct_2_std': 0.036883533000946045, 'reward_judge_quality_mean': 0.7312500476837158, 'reward_judge_quality_std': 0.3589046001434326, 'reward_total_composite_mean': 0.6683478355407715, 'reward_total_composite_std': 0.3816838264465332, 'epoch': 0.01}

 10%|▉         | 323/3300 [1:12:31<8:46:38, 10.61s/it]INFO 04-12 11:58:43 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 324/3300 [1:12:40<8:26:52, 10.22s/it]2026-04-12 11:58:52,421 | INFO | train_grpo_train | metrics_logged mode=train step=324

                                                      
{'loss': 0.1003, 'grad_norm': 14.50818920135498, 'learning_rate': 9.021212121212121e-06, 'num_tokens': 650250.0, 'completions/mean_length': 43.0, 'completions/min_length': 37.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.0, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.8761793971061707, 'rewards/meter/std': 0.18905195593833923, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5799999833106995, 'rewards/judge_quality/std': 0.3062678575515747, 'rewards/repeat_penalty/mean': 0.8346422910690308, 'rewards/repeat_penalty/std': 0.20036998391151428, 'rewards/near_duplicate_penalty/mean': 0.9579204320907593, 'rewards/near_duplicate_penalty/std': 0.03285416588187218, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9467199444770813, 'rewards/distinct_2/std': 0.07356218248605728, 'rewards/total_composite/mean': 0.5224690437316895, 'rewards/total_composite/std': 0.3168891966342926, 'reward': 0.5224690437316895, 'reward_std': 0.3168891668319702, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17705819010734558, 'sampling/sampling_logp_difference/max': 1.4733824729919434, 'sampling/importance_sampling_ratio/min': 0.22914908826351166, 'sampling/importance_sampling_ratio/mean': 1.0063377618789673, 'sampling/importance_sampling_ratio/max': 1.7909523248672485, 'entropy': 1.6156812906265259, 'clip_ratio/low_mean': 0.09179666079580784, 'clip_ratio/low_min': 0.09179666079580784, 'clip_ratio/high_mean': 0.06638756021857262, 'clip_ratio/high_max': 0.06638756021857262, 'clip_ratio/region_mean': 0.15818422101438046, 'reward_total_mean': 0.5224690437316895, 'reward_meter_mean': 0.8761793971061707, 'reward_meter_std': 0.18905195593833923, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8346422910690308, 'reward_repeat_penalty_std': 0.20036998391151428, 'reward_near_duplicate_penalty_mean': 0.9579204320907593, 'reward_near_duplicate_penalty_std': 0.03285416588187218, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9467199444770813, 'reward_distinct_2_std': 0.07356218248605728, 'reward_judge_quality_mean': 0.5799999833106995, 'reward_judge_quality_std': 0.3062678575515747, 'reward_total_composite_mean': 0.5224690437316895, 'reward_total_composite_std': 0.3168891966342926, 'epoch': 0.01}

 10%|▉         | 324/3300 [1:12:40<8:26:52, 10.22s/it]INFO 04-12 11:58:52 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 325/3300 [1:12:49<8:05:50,  9.80s/it]2026-04-12 11:59:01,234 | INFO | train_grpo_train | metrics_logged mode=train step=325

                                                      
{'loss': 0.1397, 'grad_norm': 19.418914794921875, 'learning_rate': 9.01818181818182e-06, 'num_tokens': 651765.0, 'completions/mean_length': 31.375, 'completions/min_length': 26.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.375, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.5477038025856018, 'rewards/meter/std': 0.3923529386520386, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9723190069198608, 'rewards/lexical_plausibility/std': 0.053679849952459335, 'rewards/judge_quality/mean': 0.5299999713897705, 'rewards/judge_quality/std': 0.2620250880718231, 'rewards/repeat_penalty/mean': 0.9527176022529602, 'rewards/repeat_penalty/std': 0.07624279707670212, 'rewards/near_duplicate_penalty/mean': 0.9670435190200806, 'rewards/near_duplicate_penalty/std': 0.042158085852861404, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9839743375778198, 'rewards/distinct_2/std': 0.045327357947826385, 'rewards/total_composite/mean': 0.2920708656311035, 'rewards/total_composite/std': 0.2732466459274292, 'reward': 0.2920708656311035, 'reward_std': 0.2732466459274292, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1905587911605835, 'sampling/sampling_logp_difference/max': 1.7138128280639648, 'sampling/importance_sampling_ratio/min': 0.1857045441865921, 'sampling/importance_sampling_ratio/mean': 0.9981018900871277, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0238747745752335, 'clip_ratio/low_mean': 0.0926959328353405, 'clip_ratio/low_min': 0.0926959328353405, 'clip_ratio/high_mean': 0.10422523971647024, 'clip_ratio/high_max': 0.10422523971647024, 'clip_ratio/region_mean': 0.19692117255181074, 'reward_total_mean': 0.2920708656311035, 'reward_meter_mean': 0.5477038025856018, 'reward_meter_std': 0.3923529386520386, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9723190069198608, 'reward_lexical_plausibility_std': 0.053679849952459335, 'reward_repeat_penalty_mean': 0.9527176022529602, 'reward_repeat_penalty_std': 0.07624279707670212, 'reward_near_duplicate_penalty_mean': 0.9670435190200806, 'reward_near_duplicate_penalty_std': 0.042158085852861404, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9839743375778198, 'reward_distinct_2_std': 0.045327357947826385, 'reward_judge_quality_mean': 0.5299999713897705, 'reward_judge_quality_std': 0.2620250880718231, 'reward_total_composite_mean': 0.2920708656311035, 'reward_total_composite_std': 0.2732466459274292, 'epoch': 0.01}

 10%|▉         | 325/3300 [1:12:49<8:05:50,  9.80s/it]INFO 04-12 11:59:01 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 326/3300 [1:13:00<8:33:50, 10.37s/it]2026-04-12 11:59:12,930 | INFO | train_grpo_train | metrics_logged mode=train step=326

                                                      
{'loss': 0.0514, 'grad_norm': 11.1260347366333, 'learning_rate': 9.015151515151516e-06, 'num_tokens': 654269.0, 'completions/mean_length': 106.0, 'completions/min_length': 97.0, 'completions/max_length': 117.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 106.0, 'completions/min_terminated_length': 97.0, 'completions/max_terminated_length': 117.0, 'rewards/meter/mean': 0.7726120948791504, 'rewards/meter/std': 0.20258751511573792, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.45375001430511475, 'rewards/judge_quality/std': 0.2017025649547577, 'rewards/repeat_penalty/mean': 0.9547241926193237, 'rewards/repeat_penalty/std': 0.04348671808838844, 'rewards/near_duplicate_penalty/mean': 0.9809917211532593, 'rewards/near_duplicate_penalty/std': 0.014647293835878372, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9729844331741333, 'rewards/distinct_2/std': 0.03536618500947952, 'rewards/total_composite/mean': 0.3555184602737427, 'rewards/total_composite/std': 0.2539959251880646, 'reward': 0.3555184602737427, 'reward_std': 0.25399595499038696, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17879047989845276, 'sampling/sampling_logp_difference/max': 1.7244796752929688, 'sampling/importance_sampling_ratio/min': 0.178265780210495, 'sampling/importance_sampling_ratio/mean': 1.0196211338043213, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2573939859867096, 'clip_ratio/low_mean': 0.07513099908828735, 'clip_ratio/low_min': 0.07513099908828735, 'clip_ratio/high_mean': 0.08608340192586184, 'clip_ratio/high_max': 0.08608340192586184, 'clip_ratio/region_mean': 0.1612144010141492, 'reward_total_mean': 0.3555184602737427, 'reward_meter_mean': 0.7726120948791504, 'reward_meter_std': 0.20258751511573792, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9547241926193237, 'reward_repeat_penalty_std': 0.04348671808838844, 'reward_near_duplicate_penalty_mean': 0.9809917211532593, 'reward_near_duplicate_penalty_std': 0.014647293835878372, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9729844331741333, 'reward_distinct_2_std': 0.03536618500947952, 'reward_judge_quality_mean': 0.45375001430511475, 'reward_judge_quality_std': 0.2017025649547577, 'reward_total_composite_mean': 0.3555184602737427, 'reward_total_composite_std': 0.2539959251880646, 'epoch': 0.01}

 10%|▉         | 326/3300 [1:13:00<8:33:50, 10.37s/it]INFO 04-12 11:59:13 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 327/3300 [1:13:11<8:35:33, 10.40s/it]2026-04-12 11:59:23,422 | INFO | train_grpo_train | metrics_logged mode=train step=327

                                                      
{'loss': 0.0412, 'grad_norm': 10.849813461303711, 'learning_rate': 9.012121212121213e-06, 'num_tokens': 656438.0, 'completions/mean_length': 86.125, 'completions/min_length': 70.0, 'completions/max_length': 102.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 86.125, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.46337997913360596, 'rewards/meter/std': 0.41804268956184387, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9759442806243896, 'rewards/lexical_plausibility/std': 0.06393492966890335, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.1414213478565216, 'rewards/repeat_penalty/mean': 0.9893356561660767, 'rewards/repeat_penalty/std': 0.009329397231340408, 'rewards/near_duplicate_penalty/mean': 0.9893356561660767, 'rewards/near_duplicate_penalty/std': 0.009329397231340408, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.18874692916870117, 'rewards/total_composite/std': 0.20599262416362762, 'reward': 0.18874692916870117, 'reward_std': 0.20599260926246643, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19893057644367218, 'sampling/sampling_logp_difference/max': 2.3641104698181152, 'sampling/importance_sampling_ratio/min': 0.0940329059958458, 'sampling/importance_sampling_ratio/mean': 1.0144124031066895, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7814104706048965, 'clip_ratio/low_mean': 0.10946587286889553, 'clip_ratio/low_min': 0.10946587286889553, 'clip_ratio/high_mean': 0.06408843956887722, 'clip_ratio/high_max': 0.06408843956887722, 'clip_ratio/region_mean': 0.17355431243777275, 'reward_total_mean': 0.18874692916870117, 'reward_meter_mean': 0.46337997913360596, 'reward_meter_std': 0.41804268956184387, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9759442806243896, 'reward_lexical_plausibility_std': 0.06393492966890335, 'reward_repeat_penalty_mean': 0.9893356561660767, 'reward_repeat_penalty_std': 0.009329397231340408, 'reward_near_duplicate_penalty_mean': 0.9893356561660767, 'reward_near_duplicate_penalty_std': 0.009329397231340408, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.1414213478565216, 'reward_total_composite_mean': 0.18874692916870117, 'reward_total_composite_std': 0.20599262416362762, 'epoch': 0.01}

 10%|▉         | 327/3300 [1:13:11<8:35:33, 10.40s/it]INFO 04-12 11:59:23 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 328/3300 [1:13:20<8:09:53,  9.89s/it]2026-04-12 11:59:32,111 | INFO | train_grpo_train | metrics_logged mode=train step=328

                                                      
{'loss': 0.0357, 'grad_norm': 15.33153247833252, 'learning_rate': 9.00909090909091e-06, 'num_tokens': 658067.0, 'completions/mean_length': 42.625, 'completions/min_length': 39.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.625, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.8394543528556824, 'rewards/meter/std': 0.18441522121429443, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9975490570068359, 'rewards/lexical_plausibility/std': 0.006932419259101152, 'rewards/judge_quality/mean': 0.5924999713897705, 'rewards/judge_quality/std': 0.2826785147190094, 'rewards/repeat_penalty/mean': 0.9824215173721313, 'rewards/repeat_penalty/std': 0.038053613156080246, 'rewards/near_duplicate_penalty/mean': 0.99280846118927, 'rewards/near_duplicate_penalty/std': 0.009808567352592945, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9893162250518799, 'rewards/distinct_2/std': 0.03021823801100254, 'rewards/total_composite/mean': 0.5059125423431396, 'rewards/total_composite/std': 0.2705124616622925, 'reward': 0.5059125423431396, 'reward_std': 0.2705124318599701, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20159591734409332, 'sampling/sampling_logp_difference/max': 2.097471237182617, 'sampling/importance_sampling_ratio/min': 0.12276648730039597, 'sampling/importance_sampling_ratio/mean': 1.0197023153305054, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6125789731740952, 'clip_ratio/low_mean': 0.08208791399374604, 'clip_ratio/low_min': 0.08208791399374604, 'clip_ratio/high_mean': 0.058680555783212185, 'clip_ratio/high_max': 0.058680555783212185, 'clip_ratio/region_mean': 0.14076846977695823, 'reward_total_mean': 0.5059125423431396, 'reward_meter_mean': 0.8394543528556824, 'reward_meter_std': 0.18441522121429443, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9975490570068359, 'reward_lexical_plausibility_std': 0.006932419259101152, 'reward_repeat_penalty_mean': 0.9824215173721313, 'reward_repeat_penalty_std': 0.038053613156080246, 'reward_near_duplicate_penalty_mean': 0.99280846118927, 'reward_near_duplicate_penalty_std': 0.009808567352592945, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9893162250518799, 'reward_distinct_2_std': 0.03021823801100254, 'reward_judge_quality_mean': 0.5924999713897705, 'reward_judge_quality_std': 0.2826785147190094, 'reward_total_composite_mean': 0.5059125423431396, 'reward_total_composite_std': 0.2705124616622925, 'epoch': 0.01}

 10%|▉         | 328/3300 [1:13:20<8:09:53,  9.89s/it]INFO 04-12 11:59:32 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 329/3300 [1:13:36<9:45:43, 11.83s/it]2026-04-12 11:59:48,465 | INFO | train_grpo_train | metrics_logged mode=train step=329

                                                      
{'loss': -0.0013, 'grad_norm': 20.749210357666016, 'learning_rate': 9.006060606060607e-06, 'num_tokens': 659889.0, 'completions/mean_length': 53.75, 'completions/min_length': 42.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.75, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.7675898671150208, 'rewards/meter/std': 0.28336668014526367, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.44999998807907104, 'rewards/judge_quality/std': 0.16035674512386322, 'rewards/repeat_penalty/mean': 0.9789375066757202, 'rewards/repeat_penalty/std': 0.025010893121361732, 'rewards/near_duplicate_penalty/mean': 0.9855164885520935, 'rewards/near_duplicate_penalty/std': 0.01945226453244686, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9933686852455139, 'rewards/distinct_2/std': 0.01875614933669567, 'rewards/total_composite/mean': 0.32441213726997375, 'rewards/total_composite/std': 0.19854216277599335, 'reward': 0.32441213726997375, 'reward_std': 0.19854217767715454, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19944681227207184, 'sampling/sampling_logp_difference/max': 2.3446524143218994, 'sampling/importance_sampling_ratio/min': 0.09588052332401276, 'sampling/importance_sampling_ratio/mean': 1.0091361999511719, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.358233705163002, 'clip_ratio/low_mean': 0.05262103583663702, 'clip_ratio/low_min': 0.05262103583663702, 'clip_ratio/high_mean': 0.1366305872797966, 'clip_ratio/high_max': 0.1366305872797966, 'clip_ratio/region_mean': 0.18925162311643362, 'reward_total_mean': 0.32441213726997375, 'reward_meter_mean': 0.7675898671150208, 'reward_meter_std': 0.28336668014526367, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9789375066757202, 'reward_repeat_penalty_std': 0.025010893121361732, 'reward_near_duplicate_penalty_mean': 0.9855164885520935, 'reward_near_duplicate_penalty_std': 0.01945226453244686, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9933686852455139, 'reward_distinct_2_std': 0.01875614933669567, 'reward_judge_quality_mean': 0.44999998807907104, 'reward_judge_quality_std': 0.16035674512386322, 'reward_total_composite_mean': 0.32441213726997375, 'reward_total_composite_std': 0.19854216277599335, 'epoch': 0.01}

 10%|▉         | 329/3300 [1:13:36<9:45:43, 11.83s/it]INFO 04-12 11:59:48 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 330/3300 [1:13:44<8:51:00, 10.73s/it]2026-04-12 11:59:56,622 | INFO | train_grpo_train | metrics_logged mode=train step=330

                                                      
{'loss': 0.0453, 'grad_norm': 30.89763641357422, 'learning_rate': 9.003030303030303e-06, 'num_tokens': 661208.0, 'completions/mean_length': 20.875, 'completions/min_length': 16.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.875, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.5376095771789551, 'rewards/meter/std': 0.46839410066604614, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9678649306297302, 'rewards/lexical_plausibility/std': 0.0832521840929985, 'rewards/judge_quality/mean': 0.856249988079071, 'rewards/judge_quality/std': 0.20500436425209045, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4762023091316223, 'rewards/total_composite/std': 0.45164650678634644, 'reward': 0.4762023091316223, 'reward_std': 0.4516465365886688, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18910905718803406, 'sampling/sampling_logp_difference/max': 1.443068265914917, 'sampling/importance_sampling_ratio/min': 0.2362019270658493, 'sampling/importance_sampling_ratio/mean': 1.0119171142578125, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0632958114147186, 'clip_ratio/low_mean': 0.0888278391212225, 'clip_ratio/low_min': 0.0888278391212225, 'clip_ratio/high_mean': 0.08376920688897371, 'clip_ratio/high_max': 0.08376920688897371, 'clip_ratio/region_mean': 0.1725970460101962, 'reward_total_mean': 0.4762023091316223, 'reward_meter_mean': 0.5376095771789551, 'reward_meter_std': 0.46839410066604614, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9678649306297302, 'reward_lexical_plausibility_std': 0.0832521840929985, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.856249988079071, 'reward_judge_quality_std': 0.20500436425209045, 'reward_total_composite_mean': 0.4762023091316223, 'reward_total_composite_std': 0.45164650678634644, 'epoch': 0.01}

 10%|█         | 330/3300 [1:13:44<8:51:00, 10.73s/it]INFO 04-12 11:59:56 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 331/3300 [1:13:53<8:29:23, 10.29s/it]2026-04-12 12:00:05,908 | INFO | train_grpo_train | metrics_logged mode=train step=331

                                                      
{'loss': 0.0328, 'grad_norm': 14.322080612182617, 'learning_rate': 9e-06, 'num_tokens': 662936.0, 'completions/mean_length': 51.0, 'completions/min_length': 45.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.0, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9821078777313232, 'rewards/meter/std': 0.007363711483776569, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6887500286102295, 'rewards/judge_quality/std': 0.21931307017803192, 'rewards/repeat_penalty/mean': 0.9847339987754822, 'rewards/repeat_penalty/std': 0.029361864551901817, 'rewards/near_duplicate_penalty/mean': 0.9905010461807251, 'rewards/near_duplicate_penalty/std': 0.013753361068665981, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.993990421295166, 'rewards/distinct_2/std': 0.01699776202440262, 'rewards/total_composite/mean': 0.6761473417282104, 'rewards/total_composite/std': 0.21439439058303833, 'reward': 0.6761473417282104, 'reward_std': 0.21439439058303833, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17805902659893036, 'sampling/sampling_logp_difference/max': 1.5896856784820557, 'sampling/importance_sampling_ratio/min': 0.20398969948291779, 'sampling/importance_sampling_ratio/mean': 0.988623857498169, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2775887846946716, 'clip_ratio/low_mean': 0.07430227100849152, 'clip_ratio/low_min': 0.07430227100849152, 'clip_ratio/high_mean': 0.10629522148519754, 'clip_ratio/high_max': 0.10629522148519754, 'clip_ratio/region_mean': 0.18059749249368906, 'reward_total_mean': 0.6761473417282104, 'reward_meter_mean': 0.9821078777313232, 'reward_meter_std': 0.007363711483776569, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9847339987754822, 'reward_repeat_penalty_std': 0.029361864551901817, 'reward_near_duplicate_penalty_mean': 0.9905010461807251, 'reward_near_duplicate_penalty_std': 0.013753361068665981, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.993990421295166, 'reward_distinct_2_std': 0.01699776202440262, 'reward_judge_quality_mean': 0.6887500286102295, 'reward_judge_quality_std': 0.21931307017803192, 'reward_total_composite_mean': 0.6761473417282104, 'reward_total_composite_std': 0.21439439058303833, 'epoch': 0.01}

 10%|█         | 331/3300 [1:13:53<8:29:23, 10.29s/it]INFO 04-12 12:00:06 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 332/3300 [1:14:05<8:50:02, 10.72s/it]2026-04-12 12:00:17,602 | INFO | train_grpo_train | metrics_logged mode=train step=332

                                                      
{'loss': 0.0437, 'grad_norm': 7.670846939086914, 'learning_rate': 8.996969696969697e-06, 'num_tokens': 665598.0, 'completions/mean_length': 142.75, 'completions/min_length': 125.0, 'completions/max_length': 170.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 142.75, 'completions/min_terminated_length': 125.0, 'completions/max_terminated_length': 170.0, 'rewards/meter/mean': 0.9802846908569336, 'rewards/meter/std': 0.014032049104571342, 'rewards/count_adherence/mean': 0.9642857313156128, 'rewards/count_adherence/std': 0.06613000482320786, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.09258200973272324, 'rewards/repeat_penalty/mean': 0.6013822555541992, 'rewards/repeat_penalty/std': 0.3235849440097809, 'rewards/near_duplicate_penalty/mean': 0.8782009482383728, 'rewards/near_duplicate_penalty/std': 0.10271628946065903, 'rewards/opening_diversity/mean': 0.91796875, 'rewards/opening_diversity/std': 0.09048737585544586, 'rewards/distinct_2/mean': 0.7143831253051758, 'rewards/distinct_2/std': 0.25908181071281433, 'rewards/total_composite/mean': 0.33073657751083374, 'rewards/total_composite/std': 0.09400331228971481, 'reward': 0.33073657751083374, 'reward_std': 0.09400329738855362, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16719898581504822, 'sampling/sampling_logp_difference/max': 1.2782707214355469, 'sampling/importance_sampling_ratio/min': 0.27851852774620056, 'sampling/importance_sampling_ratio/mean': 1.027553677558899, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9752332866191864, 'clip_ratio/low_mean': 0.04905758146196604, 'clip_ratio/low_min': 0.04905758146196604, 'clip_ratio/high_mean': 0.09387998469173908, 'clip_ratio/high_max': 0.09387998469173908, 'clip_ratio/region_mean': 0.14293756615370512, 'reward_total_mean': 0.33073657751083374, 'reward_meter_mean': 0.9802846908569336, 'reward_meter_std': 0.014032049104571342, 'reward_count_adherence_mean': 0.9642857313156128, 'reward_count_adherence_std': 0.06613000482320786, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6013822555541992, 'reward_repeat_penalty_std': 0.3235849440097809, 'reward_near_duplicate_penalty_mean': 0.8782009482383728, 'reward_near_duplicate_penalty_std': 0.10271628946065903, 'reward_opening_diversity_mean': 0.91796875, 'reward_opening_diversity_std': 0.09048737585544586, 'reward_distinct_2_mean': 0.7143831253051758, 'reward_distinct_2_std': 0.25908181071281433, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.09258200973272324, 'reward_total_composite_mean': 0.33073657751083374, 'reward_total_composite_std': 0.09400331228971481, 'epoch': 0.01}

 10%|█         | 332/3300 [1:14:05<8:50:02, 10.72s/it]INFO 04-12 12:00:17 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 333/3300 [1:14:14<8:22:32, 10.16s/it]2026-04-12 12:00:26,482 | INFO | train_grpo_train | metrics_logged mode=train step=333

                                                      
{'loss': 0.1484, 'grad_norm': 13.981661796569824, 'learning_rate': 8.993939393939395e-06, 'num_tokens': 667304.0, 'completions/mean_length': 47.25, 'completions/min_length': 35.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.25, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9113978147506714, 'rewards/meter/std': 0.20103241503238678, '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/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.7387499809265137, 'rewards/judge_quality/std': 0.2687770128250122, 'rewards/repeat_penalty/mean': 0.9563689231872559, 'rewards/repeat_penalty/std': 0.041896235197782516, 'rewards/near_duplicate_penalty/mean': 0.9773430824279785, 'rewards/near_duplicate_penalty/std': 0.017977163195610046, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9785662889480591, 'rewards/distinct_2/std': 0.03976903110742569, 'rewards/total_composite/mean': 0.635800838470459, 'rewards/total_composite/std': 0.30263829231262207, 'reward': 0.635800838470459, 'reward_std': 0.30263829231262207, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14866073429584503, 'sampling/sampling_logp_difference/max': 1.5793867111206055, 'sampling/importance_sampling_ratio/min': 0.2061014622449875, 'sampling/importance_sampling_ratio/mean': 1.0088062286376953, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1610790491104126, 'clip_ratio/low_mean': 0.06341799814254045, 'clip_ratio/low_min': 0.06341799814254045, 'clip_ratio/high_mean': 0.06228313175961375, 'clip_ratio/high_max': 0.06228313175961375, 'clip_ratio/region_mean': 0.1257011299021542, 'reward_total_mean': 0.635800838470459, 'reward_meter_mean': 0.9113978147506714, 'reward_meter_std': 0.20103241503238678, '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_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.9563689231872559, 'reward_repeat_penalty_std': 0.041896235197782516, 'reward_near_duplicate_penalty_mean': 0.9773430824279785, 'reward_near_duplicate_penalty_std': 0.017977163195610046, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9785662889480591, 'reward_distinct_2_std': 0.03976903110742569, 'reward_judge_quality_mean': 0.7387499809265137, 'reward_judge_quality_std': 0.2687770128250122, 'reward_total_composite_mean': 0.635800838470459, 'reward_total_composite_std': 0.30263829231262207, 'epoch': 0.01}

 10%|█         | 333/3300 [1:14:14<8:22:32, 10.16s/it]INFO 04-12 12:00:26 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 334/3300 [1:14:32<10:24:56, 12.64s/it]2026-04-12 12:00:44,906 | INFO | train_grpo_train | metrics_logged mode=train step=334

                                                       
{'loss': -0.1468, 'grad_norm': 4.256762981414795, 'learning_rate': 8.990909090909092e-06, 'num_tokens': 670580.0, 'completions/mean_length': 249.5, 'completions/min_length': 194.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 212.00001525878906, 'completions/min_terminated_length': 194.0, 'completions/max_terminated_length': 249.0, 'rewards/meter/mean': 0.8910650014877319, 'rewards/meter/std': 0.2547892928123474, 'rewards/count_adherence/mean': 0.8624999523162842, 'rewards/count_adherence/std': 0.07440237700939178, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9741270542144775, 'rewards/lexical_plausibility/std': 0.07317972183227539, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.1505940556526184, 'rewards/repeat_penalty/mean': 0.7558687329292297, 'rewards/repeat_penalty/std': 0.299955278635025, 'rewards/near_duplicate_penalty/mean': 0.9519389867782593, 'rewards/near_duplicate_penalty/std': 0.05630122125148773, 'rewards/opening_diversity/mean': 0.9261363744735718, 'rewards/opening_diversity/std': 0.20891793072223663, 'rewards/distinct_2/mean': 0.8294985294342041, 'rewards/distinct_2/std': 0.19569948315620422, 'rewards/total_composite/mean': 0.2373369038105011, 'rewards/total_composite/std': 0.13343991339206696, 'reward': 0.2373369038105011, 'reward_std': 0.13343991339206696, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18489407002925873, 'sampling/sampling_logp_difference/max': 1.5506134033203125, 'sampling/importance_sampling_ratio/min': 0.21211782097816467, 'sampling/importance_sampling_ratio/mean': 1.037339687347412, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7812134623527527, 'clip_ratio/low_mean': 0.03522298112511635, 'clip_ratio/low_min': 0.03522298112511635, 'clip_ratio/high_mean': 0.09229760803282261, 'clip_ratio/high_max': 0.09229760803282261, 'clip_ratio/region_mean': 0.12752058915793896, 'reward_total_mean': 0.2373369038105011, 'reward_meter_mean': 0.8910650014877319, 'reward_meter_std': 0.2547892928123474, 'reward_count_adherence_mean': 0.8624999523162842, 'reward_count_adherence_std': 0.07440237700939178, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9741270542144775, 'reward_lexical_plausibility_std': 0.07317972183227539, 'reward_repeat_penalty_mean': 0.7558687329292297, 'reward_repeat_penalty_std': 0.299955278635025, 'reward_near_duplicate_penalty_mean': 0.9519389867782593, 'reward_near_duplicate_penalty_std': 0.05630122125148773, 'reward_opening_diversity_mean': 0.9261363744735718, 'reward_opening_diversity_std': 0.20891793072223663, 'reward_distinct_2_mean': 0.8294985294342041, 'reward_distinct_2_std': 0.19569948315620422, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.1505940556526184, 'reward_total_composite_mean': 0.2373369038105011, 'reward_total_composite_std': 0.13343991339206696, 'epoch': 0.01}

 10%|█         | 334/3300 [1:14:32<10:24:56, 12.64s/it]INFO 04-12 12:00:45 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 335/3300 [1:14:43<9:52:52, 12.00s/it] 2026-04-12 12:00:55,398 | INFO | train_grpo_train | metrics_logged mode=train step=335

                                                      
{'loss': -0.03, 'grad_norm': 20.10825538635254, 'learning_rate': 8.98787878787879e-06, 'num_tokens': 672044.0, 'completions/mean_length': 23.0, 'completions/min_length': 16.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.0, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.5148899555206299, 'rewards/meter/std': 0.41198641061782837, '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/lexical_plausibility/mean': 0.7461690902709961, 'rewards/lexical_plausibility/std': 0.26617443561553955, 'rewards/judge_quality/mean': 0.7300000190734863, 'rewards/judge_quality/std': 0.3899816870689392, 'rewards/repeat_penalty/mean': 0.78125, 'rewards/repeat_penalty/std': 0.0883883461356163, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.40092235803604126, 'rewards/total_composite/std': 0.382964164018631, 'reward': 0.40092235803604126, 'reward_std': 0.3829641342163086, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.190692737698555, 'sampling/sampling_logp_difference/max': 1.3405847549438477, 'sampling/importance_sampling_ratio/min': 0.261692613363266, 'sampling/importance_sampling_ratio/mean': 1.0392323732376099, 'sampling/importance_sampling_ratio/max': 1.9872056245803833, 'entropy': 2.044335350394249, 'clip_ratio/low_mean': 0.09232696797698736, 'clip_ratio/low_min': 0.09232696797698736, 'clip_ratio/high_mean': 0.07881728745996952, 'clip_ratio/high_max': 0.07881728745996952, 'clip_ratio/region_mean': 0.17114425543695688, 'reward_total_mean': 0.40092235803604126, 'reward_meter_mean': 0.5148899555206299, 'reward_meter_std': 0.41198641061782837, '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_lexical_plausibility_mean': 0.7461690902709961, 'reward_lexical_plausibility_std': 0.26617443561553955, 'reward_repeat_penalty_mean': 0.78125, 'reward_repeat_penalty_std': 0.0883883461356163, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7300000190734863, 'reward_judge_quality_std': 0.3899816870689392, 'reward_total_composite_mean': 0.40092235803604126, 'reward_total_composite_std': 0.382964164018631, 'epoch': 0.01}

 10%|█         | 335/3300 [1:14:43<9:52:52, 12.00s/it]INFO 04-12 12:00:55 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 336/3300 [1:14:55<9:54:48, 12.04s/it]2026-04-12 12:01:07,539 | INFO | train_grpo_train | metrics_logged mode=train step=336

                                                      
{'loss': 0.0497, 'grad_norm': 11.71253490447998, 'learning_rate': 8.984848484848485e-06, 'num_tokens': 673663.0, 'completions/mean_length': 38.375, 'completions/min_length': 32.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.375, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.4177127480506897, 'rewards/meter/std': 0.2962014079093933, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7975000143051147, 'rewards/judge_quality/std': 0.2471408098936081, 'rewards/repeat_penalty/mean': 0.965746283531189, 'rewards/repeat_penalty/std': 0.09504672139883041, 'rewards/near_duplicate_penalty/mean': 0.9845749139785767, 'rewards/near_duplicate_penalty/std': 0.041808515787124634, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9786324501037598, 'rewards/distinct_2/std': 0.06043647602200508, 'rewards/total_composite/mean': 0.31122303009033203, 'rewards/total_composite/std': 0.23601660132408142, 'reward': 0.31122303009033203, 'reward_std': 0.23601660132408142, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15965794026851654, 'sampling/sampling_logp_difference/max': 1.514892578125, 'sampling/importance_sampling_ratio/min': 0.21983179450035095, 'sampling/importance_sampling_ratio/mean': 1.0127167701721191, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7997074276208878, 'clip_ratio/low_mean': 0.08659902401268482, 'clip_ratio/low_min': 0.08659902401268482, 'clip_ratio/high_mean': 0.05541834328323603, 'clip_ratio/high_max': 0.05541834328323603, 'clip_ratio/region_mean': 0.14201736729592085, 'reward_total_mean': 0.31122303009033203, 'reward_meter_mean': 0.4177127480506897, 'reward_meter_std': 0.2962014079093933, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.965746283531189, 'reward_repeat_penalty_std': 0.09504672139883041, 'reward_near_duplicate_penalty_mean': 0.9845749139785767, 'reward_near_duplicate_penalty_std': 0.041808515787124634, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9786324501037598, 'reward_distinct_2_std': 0.06043647602200508, 'reward_judge_quality_mean': 0.7975000143051147, 'reward_judge_quality_std': 0.2471408098936081, 'reward_total_composite_mean': 0.31122303009033203, 'reward_total_composite_std': 0.23601660132408142, 'epoch': 0.01}

 10%|█         | 336/3300 [1:14:55<9:54:48, 12.04s/it]INFO 04-12 12:01:07 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 337/3300 [1:15:05<9:19:43, 11.33s/it]2026-04-12 12:01:17,226 | INFO | train_grpo_train | metrics_logged mode=train step=337

                                                      
{'loss': 0.1152, 'grad_norm': 16.95784568786621, 'learning_rate': 8.981818181818182e-06, 'num_tokens': 675144.0, 'completions/mean_length': 35.125, 'completions/min_length': 28.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.125, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.49531427025794983, 'rewards/meter/std': 0.41725823283195496, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5637499690055847, 'rewards/judge_quality/std': 0.3091896176338196, 'rewards/repeat_penalty/mean': 0.9841508865356445, 'rewards/repeat_penalty/std': 0.019338395446538925, 'rewards/near_duplicate_penalty/mean': 0.9841508865356445, 'rewards/near_duplicate_penalty/std': 0.019338395446538925, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3064648509025574, 'rewards/total_composite/std': 0.3691054582595825, 'reward': 0.3064648509025574, 'reward_std': 0.3691054582595825, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19438858330249786, 'sampling/sampling_logp_difference/max': 2.089127540588379, 'sampling/importance_sampling_ratio/min': 0.12379509955644608, 'sampling/importance_sampling_ratio/mean': 0.991215705871582, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0955918431282043, 'clip_ratio/low_mean': 0.12712144292891026, 'clip_ratio/low_min': 0.12712144292891026, 'clip_ratio/high_mean': 0.04520089365541935, 'clip_ratio/high_max': 0.04520089365541935, 'clip_ratio/region_mean': 0.1723223365843296, 'reward_total_mean': 0.3064648509025574, 'reward_meter_mean': 0.49531427025794983, 'reward_meter_std': 0.41725823283195496, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9841508865356445, 'reward_repeat_penalty_std': 0.019338395446538925, 'reward_near_duplicate_penalty_mean': 0.9841508865356445, 'reward_near_duplicate_penalty_std': 0.019338395446538925, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5637499690055847, 'reward_judge_quality_std': 0.3091896176338196, 'reward_total_composite_mean': 0.3064648509025574, 'reward_total_composite_std': 0.3691054582595825, 'epoch': 0.01}

 10%|█         | 337/3300 [1:15:05<9:19:43, 11.33s/it]INFO 04-12 12:01:17 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 338/3300 [1:15:14<8:45:19, 10.64s/it]2026-04-12 12:01:26,253 | INFO | train_grpo_train | metrics_logged mode=train step=338

                                                      
{'loss': -0.0405, 'grad_norm': 16.093406677246094, 'learning_rate': 8.97878787878788e-06, 'num_tokens': 676659.0, 'completions/mean_length': 35.375, 'completions/min_length': 30.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.375, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.8954558372497559, 'rewards/meter/std': 0.23072436451911926, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6850000023841858, 'rewards/judge_quality/std': 0.2512255907058716, 'rewards/repeat_penalty/mean': 0.8931442499160767, 'rewards/repeat_penalty/std': 0.15029844641685486, 'rewards/near_duplicate_penalty/mean': 0.9661966562271118, 'rewards/near_duplicate_penalty/std': 0.038574304431676865, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9477058053016663, 'rewards/distinct_2/std': 0.07641414552927017, 'rewards/total_composite/mean': 0.6326240301132202, 'rewards/total_composite/std': 0.30069440603256226, 'reward': 0.6326240301132202, 'reward_std': 0.30069440603256226, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1786806881427765, 'sampling/sampling_logp_difference/max': 1.7570123672485352, 'sampling/importance_sampling_ratio/min': 0.17255963385105133, 'sampling/importance_sampling_ratio/mean': 1.0094597339630127, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.328508049249649, 'clip_ratio/low_mean': 0.08336466364562511, 'clip_ratio/low_min': 0.08336466364562511, 'clip_ratio/high_mean': 0.12179983593523502, 'clip_ratio/high_max': 0.12179983593523502, 'clip_ratio/region_mean': 0.20516449958086014, 'reward_total_mean': 0.6326240301132202, 'reward_meter_mean': 0.8954558372497559, 'reward_meter_std': 0.23072436451911926, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8931442499160767, 'reward_repeat_penalty_std': 0.15029844641685486, 'reward_near_duplicate_penalty_mean': 0.9661966562271118, 'reward_near_duplicate_penalty_std': 0.038574304431676865, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9477058053016663, 'reward_distinct_2_std': 0.07641414552927017, 'reward_judge_quality_mean': 0.6850000023841858, 'reward_judge_quality_std': 0.2512255907058716, 'reward_total_composite_mean': 0.6326240301132202, 'reward_total_composite_std': 0.30069440603256226, 'epoch': 0.01}

 10%|█         | 338/3300 [1:15:14<8:45:19, 10.64s/it]INFO 04-12 12:01:26 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 339/3300 [1:15:24<8:45:10, 10.64s/it]2026-04-12 12:01:36,899 | INFO | train_grpo_train | metrics_logged mode=train step=339

                                                      
{'loss': -0.0032, 'grad_norm': 8.856263160705566, 'learning_rate': 8.975757575757577e-06, 'num_tokens': 679299.0, 'completions/mean_length': 127.0, 'completions/min_length': 117.0, 'completions/max_length': 136.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 127.0, 'completions/min_terminated_length': 117.0, 'completions/max_terminated_length': 136.0, 'rewards/meter/mean': 0.9636448621749878, 'rewards/meter/std': 0.0717000663280487, 'rewards/count_adherence/mean': 0.9791666269302368, 'rewards/count_adherence/std': 0.0589255727827549, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.4897316098213196, 'rewards/repeat_penalty/std': 0.3254680037498474, 'rewards/near_duplicate_penalty/mean': 0.8693605065345764, 'rewards/near_duplicate_penalty/std': 0.0926419049501419, 'rewards/opening_diversity/mean': 0.9603365659713745, 'rewards/opening_diversity/std': 0.07561343163251877, 'rewards/distinct_2/mean': 0.5852110385894775, 'rewards/distinct_2/std': 0.2783803343772888, 'rewards/total_composite/mean': 0.25445759296417236, 'rewards/total_composite/std': 0.14898806810379028, 'reward': 0.25445759296417236, 'reward_std': 0.1489880532026291, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1530618965625763, 'sampling/sampling_logp_difference/max': 1.951028823852539, 'sampling/importance_sampling_ratio/min': 0.14212776720523834, 'sampling/importance_sampling_ratio/mean': 1.027160406112671, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3738272562623024, 'clip_ratio/low_mean': 0.04581594280898571, 'clip_ratio/low_min': 0.04581594280898571, 'clip_ratio/high_mean': 0.09792396426200867, 'clip_ratio/high_max': 0.09792396426200867, 'clip_ratio/region_mean': 0.14373990707099438, 'reward_total_mean': 0.25445759296417236, 'reward_meter_mean': 0.9636448621749878, 'reward_meter_std': 0.0717000663280487, 'reward_count_adherence_mean': 0.9791666269302368, 'reward_count_adherence_std': 0.0589255727827549, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.4897316098213196, 'reward_repeat_penalty_std': 0.3254680037498474, 'reward_near_duplicate_penalty_mean': 0.8693605065345764, 'reward_near_duplicate_penalty_std': 0.0926419049501419, 'reward_opening_diversity_mean': 0.9603365659713745, 'reward_opening_diversity_std': 0.07561343163251877, 'reward_distinct_2_mean': 0.5852110385894775, 'reward_distinct_2_std': 0.2783803343772888, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.25445759296417236, 'reward_total_composite_std': 0.14898806810379028, 'epoch': 0.01}

 10%|█         | 339/3300 [1:15:24<8:45:10, 10.64s/it]INFO 04-12 12:01:37 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 340/3300 [1:15:36<8:52:51, 10.80s/it]2026-04-12 12:01:48,078 | INFO | train_grpo_train | metrics_logged mode=train step=340

                                                      
{'loss': 0.1851, 'grad_norm': 10.916351318359375, 'learning_rate': 8.972727272727272e-06, 'num_tokens': 681516.0, 'completions/mean_length': 96.125, 'completions/min_length': 82.0, 'completions/max_length': 139.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 96.125, 'completions/min_terminated_length': 82.0, 'completions/max_terminated_length': 139.0, 'rewards/meter/mean': 0.8985081911087036, 'rewards/meter/std': 0.16623960435390472, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.05345224589109421, 'rewards/repeat_penalty/mean': 0.8063184022903442, 'rewards/repeat_penalty/std': 0.28554925322532654, 'rewards/near_duplicate_penalty/mean': 0.9560104608535767, 'rewards/near_duplicate_penalty/std': 0.04691820591688156, 'rewards/opening_diversity/mean': 0.949999988079071, 'rewards/opening_diversity/std': 0.1414213478565216, 'rewards/distinct_2/mean': 0.8740012049674988, 'rewards/distinct_2/std': 0.20042245090007782, 'rewards/total_composite/mean': 0.3324708938598633, 'rewards/total_composite/std': 0.1486913114786148, 'reward': 0.3324708938598633, 'reward_std': 0.1486913114786148, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19740067422389984, 'sampling/sampling_logp_difference/max': 1.6764140129089355, 'sampling/importance_sampling_ratio/min': 0.18704351782798767, 'sampling/importance_sampling_ratio/mean': 1.018814206123352, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8450011909008026, 'clip_ratio/low_mean': 0.055041538551449776, 'clip_ratio/low_min': 0.055041538551449776, 'clip_ratio/high_mean': 0.15035482496023178, 'clip_ratio/high_max': 0.15035482496023178, 'clip_ratio/region_mean': 0.20539636351168156, 'reward_total_mean': 0.3324708938598633, 'reward_meter_mean': 0.8985081911087036, 'reward_meter_std': 0.16623960435390472, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.09258200973272324, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8063184022903442, 'reward_repeat_penalty_std': 0.28554925322532654, 'reward_near_duplicate_penalty_mean': 0.9560104608535767, 'reward_near_duplicate_penalty_std': 0.04691820591688156, 'reward_opening_diversity_mean': 0.949999988079071, 'reward_opening_diversity_std': 0.1414213478565216, 'reward_distinct_2_mean': 0.8740012049674988, 'reward_distinct_2_std': 0.20042245090007782, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.05345224589109421, 'reward_total_composite_mean': 0.3324708938598633, 'reward_total_composite_std': 0.1486913114786148, 'epoch': 0.01}

 10%|█         | 340/3300 [1:15:36<8:52:51, 10.80s/it]INFO 04-12 12:01:48 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 341/3300 [1:15:53<10:27:15, 12.72s/it]2026-04-12 12:02:05,259 | INFO | train_grpo_train | metrics_logged mode=train step=341

                                                       
{'loss': -0.1113, 'grad_norm': 3.631061315536499, 'learning_rate': 8.969696969696971e-06, 'num_tokens': 685947.0, 'completions/mean_length': 359.875, 'completions/min_length': 280.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 338.14288330078125, 'completions/min_terminated_length': 280.0, 'completions/max_terminated_length': 415.0, 'rewards/meter/mean': 0.7854980230331421, 'rewards/meter/std': 0.36010047793388367, 'rewards/count_adherence/mean': 0.7980769276618958, 'rewards/count_adherence/std': 0.2134033739566803, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9390830993652344, 'rewards/lexical_plausibility/std': 0.17229899764060974, 'rewards/judge_quality/mean': 0.21250000596046448, 'rewards/judge_quality/std': 0.13024701178073883, 'rewards/repeat_penalty/mean': 0.3892909288406372, 'rewards/repeat_penalty/std': 0.4289466440677643, 'rewards/near_duplicate_penalty/mean': 0.7990989685058594, 'rewards/near_duplicate_penalty/std': 0.19495080411434174, 'rewards/opening_diversity/mean': 0.8870849013328552, 'rewards/opening_diversity/std': 0.20291320979595184, 'rewards/distinct_2/mean': 0.4373234510421753, 'rewards/distinct_2/std': 0.424831748008728, 'rewards/total_composite/mean': 0.11135604977607727, 'rewards/total_composite/std': 0.10392560809850693, 'reward': 0.11135604977607727, 'reward_std': 0.10392560809850693, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10169584304094315, 'sampling/sampling_logp_difference/max': 1.721909999847412, 'sampling/importance_sampling_ratio/min': 0.178724467754364, 'sampling/importance_sampling_ratio/mean': 1.0194363594055176, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9587608352303505, 'clip_ratio/low_mean': 0.02962994109839201, 'clip_ratio/low_min': 0.02962994109839201, 'clip_ratio/high_mean': 0.059364788699895144, 'clip_ratio/high_max': 0.059364788699895144, 'clip_ratio/region_mean': 0.08899472979828715, 'reward_total_mean': 0.11135604977607727, 'reward_meter_mean': 0.7854980230331421, 'reward_meter_std': 0.36010047793388367, 'reward_count_adherence_mean': 0.7980769276618958, 'reward_count_adherence_std': 0.2134033739566803, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9390830993652344, 'reward_lexical_plausibility_std': 0.17229899764060974, 'reward_repeat_penalty_mean': 0.3892909288406372, 'reward_repeat_penalty_std': 0.4289466440677643, 'reward_near_duplicate_penalty_mean': 0.7990989685058594, 'reward_near_duplicate_penalty_std': 0.19495080411434174, 'reward_opening_diversity_mean': 0.8870849013328552, 'reward_opening_diversity_std': 0.20291320979595184, 'reward_distinct_2_mean': 0.4373234510421753, 'reward_distinct_2_std': 0.424831748008728, 'reward_judge_quality_mean': 0.21250000596046448, 'reward_judge_quality_std': 0.13024701178073883, 'reward_total_composite_mean': 0.11135604977607727, 'reward_total_composite_std': 0.10392560809850693, 'epoch': 0.01}

 10%|█         | 341/3300 [1:15:53<10:27:15, 12.72s/it]INFO 04-12 12:02:05 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 342/3300 [1:16:01<9:26:03, 11.48s/it] 2026-04-12 12:02:13,855 | INFO | train_grpo_train | metrics_logged mode=train step=342

                                                      
{'loss': -0.0229, 'grad_norm': 12.50191593170166, 'learning_rate': 8.966666666666667e-06, 'num_tokens': 687799.0, 'completions/mean_length': 61.5, 'completions/min_length': 46.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.5, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.29511725902557373, 'rewards/meter/std': 0.3623807728290558, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9977971315383911, 'rewards/lexical_plausibility/std': 0.006230758503079414, 'rewards/judge_quality/mean': 0.5425000190734863, 'rewards/judge_quality/std': 0.23705033957958221, 'rewards/repeat_penalty/mean': 0.9774243235588074, 'rewards/repeat_penalty/std': 0.020903760567307472, 'rewards/near_duplicate_penalty/mean': 0.9774243235588074, 'rewards/near_duplicate_penalty/std': 0.020903760567307472, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.14046093821525574, 'rewards/total_composite/std': 0.13782621920108795, 'reward': 0.14046093821525574, 'reward_std': 0.13782620429992676, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18585516512393951, 'sampling/sampling_logp_difference/max': 1.7113144397735596, 'sampling/importance_sampling_ratio/min': 0.1806282103061676, 'sampling/importance_sampling_ratio/mean': 1.0404918193817139, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7546063512563705, 'clip_ratio/low_mean': 0.10133026633411646, 'clip_ratio/low_min': 0.10133026633411646, 'clip_ratio/high_mean': 0.0722538661211729, 'clip_ratio/high_max': 0.0722538661211729, 'clip_ratio/region_mean': 0.17358413245528936, 'reward_total_mean': 0.14046093821525574, 'reward_meter_mean': 0.29511725902557373, 'reward_meter_std': 0.3623807728290558, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9977971315383911, 'reward_lexical_plausibility_std': 0.006230758503079414, 'reward_repeat_penalty_mean': 0.9774243235588074, 'reward_repeat_penalty_std': 0.020903760567307472, 'reward_near_duplicate_penalty_mean': 0.9774243235588074, 'reward_near_duplicate_penalty_std': 0.020903760567307472, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5425000190734863, 'reward_judge_quality_std': 0.23705033957958221, 'reward_total_composite_mean': 0.14046093821525574, 'reward_total_composite_std': 0.13782621920108795, 'epoch': 0.01}

 10%|█         | 342/3300 [1:16:01<9:26:03, 11.48s/it]INFO 04-12 12:02:14 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 343/3300 [1:16:15<10:01:45, 12.21s/it]2026-04-12 12:02:27,769 | INFO | train_grpo_train | metrics_logged mode=train step=343

                                                       
{'loss': -0.0237, 'grad_norm': 5.930196762084961, 'learning_rate': 8.963636363636364e-06, 'num_tokens': 689356.0, 'completions/mean_length': 96.625, 'completions/min_length': 31.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 37.28571701049805, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.3506585359573364, 'rewards/meter/std': 0.4087865650653839, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.9431853294372559, 'rewards/lexical_plausibility/std': 0.1606961339712143, 'rewards/judge_quality/mean': 0.3400000035762787, 'rewards/judge_quality/std': 0.2768702805042267, 'rewards/repeat_penalty/mean': 0.9897685647010803, 'rewards/repeat_penalty/std': 0.014945391565561295, 'rewards/near_duplicate_penalty/mean': 0.9897685647010803, 'rewards/near_duplicate_penalty/std': 0.014945391565561295, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1779477298259735, 'rewards/total_composite/std': 0.3148679733276367, 'reward': 0.1779477298259735, 'reward_std': 0.31486794352531433, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21920634806156158, 'sampling/sampling_logp_difference/max': 1.681375503540039, 'sampling/importance_sampling_ratio/min': 0.18611779808998108, 'sampling/importance_sampling_ratio/mean': 1.0393905639648438, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.838468760251999, 'clip_ratio/low_mean': 0.11618281900882721, 'clip_ratio/low_min': 0.11618281900882721, 'clip_ratio/high_mean': 0.0432076808065176, 'clip_ratio/high_max': 0.0432076808065176, 'clip_ratio/region_mean': 0.1593904998153448, 'reward_total_mean': 0.1779477298259735, 'reward_meter_mean': 0.3506585359573364, 'reward_meter_std': 0.4087865650653839, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.9431853294372559, 'reward_lexical_plausibility_std': 0.1606961339712143, 'reward_repeat_penalty_mean': 0.9897685647010803, 'reward_repeat_penalty_std': 0.014945391565561295, 'reward_near_duplicate_penalty_mean': 0.9897685647010803, 'reward_near_duplicate_penalty_std': 0.014945391565561295, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3400000035762787, 'reward_judge_quality_std': 0.2768702805042267, 'reward_total_composite_mean': 0.1779477298259735, 'reward_total_composite_std': 0.3148679733276367, 'epoch': 0.01}

 10%|█         | 343/3300 [1:16:15<10:01:45, 12.21s/it]INFO 04-12 12:02:28 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 344/3300 [1:16:23<8:56:50, 10.90s/it] 2026-04-12 12:02:35,595 | INFO | train_grpo_train | metrics_logged mode=train step=344

                                                      
{'loss': 0.0856, 'grad_norm': 28.26715087890625, 'learning_rate': 8.960606060606061e-06, 'num_tokens': 690706.0, 'completions/mean_length': 22.75, 'completions/min_length': 20.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.75, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.5608992576599121, 'rewards/meter/std': 0.42065587639808655, '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/lexical_plausibility/mean': 0.8341332674026489, 'rewards/lexical_plausibility/std': 0.16632896661758423, 'rewards/judge_quality/mean': 0.8350000381469727, 'rewards/judge_quality/std': 0.27717968821525574, 'rewards/repeat_penalty/mean': 0.7769063115119934, 'rewards/repeat_penalty/std': 0.1433175951242447, 'rewards/near_duplicate_penalty/mean': 0.952541708946228, 'rewards/near_duplicate_penalty/std': 0.06054948642849922, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.49674567580223083, 'rewards/total_composite/std': 0.41812607645988464, 'reward': 0.49674567580223083, 'reward_std': 0.41812607645988464, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2103598564863205, 'sampling/sampling_logp_difference/max': 1.5652546882629395, 'sampling/importance_sampling_ratio/min': 0.2090347558259964, 'sampling/importance_sampling_ratio/mean': 1.0092546939849854, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.388786643743515, 'clip_ratio/low_mean': 0.07365959882736206, 'clip_ratio/low_min': 0.07365959882736206, 'clip_ratio/high_mean': 0.09747023973613977, 'clip_ratio/high_max': 0.09747023973613977, 'clip_ratio/region_mean': 0.17112983856350183, 'reward_total_mean': 0.49674567580223083, 'reward_meter_mean': 0.5608992576599121, 'reward_meter_std': 0.42065587639808655, '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_lexical_plausibility_mean': 0.8341332674026489, 'reward_lexical_plausibility_std': 0.16632896661758423, 'reward_repeat_penalty_mean': 0.7769063115119934, 'reward_repeat_penalty_std': 0.1433175951242447, 'reward_near_duplicate_penalty_mean': 0.952541708946228, 'reward_near_duplicate_penalty_std': 0.06054948642849922, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8350000381469727, 'reward_judge_quality_std': 0.27717968821525574, 'reward_total_composite_mean': 0.49674567580223083, 'reward_total_composite_std': 0.41812607645988464, 'epoch': 0.01}

 10%|█         | 344/3300 [1:16:23<8:56:50, 10.90s/it]INFO 04-12 12:02:35 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 345/3300 [1:16:38<9:58:35, 12.15s/it]2026-04-12 12:02:50,687 | INFO | train_grpo_train | metrics_logged mode=train step=345

                                                      
{'loss': -0.0807, 'grad_norm': 1.9718246459960938, 'learning_rate': 8.957575757575758e-06, 'num_tokens': 694508.0, 'completions/mean_length': 478.25, 'completions/min_length': 303.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 444.5, 'completions/min_terminated_length': 303.0, 'completions/max_terminated_length': 511.0, 'rewards/meter/mean': 0.8270421624183655, 'rewards/meter/std': 0.27804070711135864, 'rewards/count_adherence/mean': 0.6312499642372131, 'rewards/count_adherence/std': 0.18310320377349854, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9499764442443848, 'rewards/lexical_plausibility/std': 0.09777331352233887, 'rewards/judge_quality/mean': 0.125, 'rewards/judge_quality/std': 0.046291008591651917, 'rewards/repeat_penalty/mean': 0.26956379413604736, 'rewards/repeat_penalty/std': 0.44275808334350586, 'rewards/near_duplicate_penalty/mean': 0.7016512751579285, 'rewards/near_duplicate_penalty/std': 0.22374014556407928, 'rewards/opening_diversity/mean': 0.6996302008628845, 'rewards/opening_diversity/std': 0.3217163383960724, 'rewards/distinct_2/mean': 0.3010595738887787, 'rewards/distinct_2/std': 0.4387643337249756, 'rewards/total_composite/mean': 0.07686956971883774, 'rewards/total_composite/std': 0.049842629581689835, 'reward': 0.07686956971883774, 'reward_std': 0.049842629581689835, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12953819334506989, 'sampling/sampling_logp_difference/max': 1.0903377532958984, 'sampling/importance_sampling_ratio/min': 0.3361029624938965, 'sampling/importance_sampling_ratio/mean': 1.0332744121551514, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.354970782995224, 'clip_ratio/low_mean': 0.013613861054182053, 'clip_ratio/low_min': 0.013613861054182053, 'clip_ratio/high_mean': 0.037701843306422234, 'clip_ratio/high_max': 0.037701843306422234, 'clip_ratio/region_mean': 0.051315704360604286, 'reward_total_mean': 0.07686956971883774, 'reward_meter_mean': 0.8270421624183655, 'reward_meter_std': 0.27804070711135864, 'reward_count_adherence_mean': 0.6312499642372131, 'reward_count_adherence_std': 0.18310320377349854, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9499764442443848, 'reward_lexical_plausibility_std': 0.09777331352233887, 'reward_repeat_penalty_mean': 0.26956379413604736, 'reward_repeat_penalty_std': 0.44275808334350586, 'reward_near_duplicate_penalty_mean': 0.7016512751579285, 'reward_near_duplicate_penalty_std': 0.22374014556407928, 'reward_opening_diversity_mean': 0.6996302008628845, 'reward_opening_diversity_std': 0.3217163383960724, 'reward_distinct_2_mean': 0.3010595738887787, 'reward_distinct_2_std': 0.4387643337249756, 'reward_judge_quality_mean': 0.125, 'reward_judge_quality_std': 0.046291008591651917, 'reward_total_composite_mean': 0.07686956971883774, 'reward_total_composite_std': 0.049842629581689835, 'epoch': 0.01}

 10%|█         | 345/3300 [1:16:38<9:58:35, 12.15s/it]INFO 04-12 12:02:50 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 346/3300 [1:16:48<9:25:46, 11.49s/it]2026-04-12 12:03:00,631 | INFO | train_grpo_train | metrics_logged mode=train step=346

                                                      
{'loss': 0.0856, 'grad_norm': 19.72292709350586, 'learning_rate': 8.954545454545456e-06, 'num_tokens': 696232.0, 'completions/mean_length': 32.5, 'completions/min_length': 22.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.5, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.493013471364975, 'rewards/meter/std': 0.4109993875026703, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9787946343421936, 'rewards/lexical_plausibility/std': 0.04103745147585869, 'rewards/judge_quality/mean': 0.8100000023841858, 'rewards/judge_quality/std': 0.22258225083351135, 'rewards/repeat_penalty/mean': 0.9584251642227173, 'rewards/repeat_penalty/std': 0.03874953091144562, 'rewards/near_duplicate_penalty/mean': 0.9584251642227173, 'rewards/near_duplicate_penalty/std': 0.03874953091144562, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.38943254947662354, 'rewards/total_composite/std': 0.3704363703727722, 'reward': 0.38943254947662354, 'reward_std': 0.3704363703727722, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18848317861557007, 'sampling/sampling_logp_difference/max': 1.496908187866211, 'sampling/importance_sampling_ratio/min': 0.22382111847400665, 'sampling/importance_sampling_ratio/mean': 0.9938216209411621, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9225008711218834, 'clip_ratio/low_mean': 0.10501706041395664, 'clip_ratio/low_min': 0.10501706041395664, 'clip_ratio/high_mean': 0.08121081627905369, 'clip_ratio/high_max': 0.08121081627905369, 'clip_ratio/region_mean': 0.18622787669301033, 'reward_total_mean': 0.38943254947662354, 'reward_meter_mean': 0.493013471364975, 'reward_meter_std': 0.4109993875026703, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9787946343421936, 'reward_lexical_plausibility_std': 0.04103745147585869, 'reward_repeat_penalty_mean': 0.9584251642227173, 'reward_repeat_penalty_std': 0.03874953091144562, 'reward_near_duplicate_penalty_mean': 0.9584251642227173, 'reward_near_duplicate_penalty_std': 0.03874953091144562, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8100000023841858, 'reward_judge_quality_std': 0.22258225083351135, 'reward_total_composite_mean': 0.38943254947662354, 'reward_total_composite_std': 0.3704363703727722, 'epoch': 0.01}

 10%|█         | 346/3300 [1:16:48<9:25:46, 11.49s/it]INFO 04-12 12:03:00 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 347/3300 [1:16:57<8:41:28, 10.60s/it]2026-04-12 12:03:09,135 | INFO | train_grpo_train | metrics_logged mode=train step=347

                                                      
{'loss': -0.0015, 'grad_norm': 19.128320693969727, 'learning_rate': 8.951515151515153e-06, 'num_tokens': 697917.0, 'completions/mean_length': 36.625, 'completions/min_length': 35.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.625, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.9164797067642212, 'rewards/meter/std': 0.021714089438319206, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8487499952316284, 'rewards/judge_quality/std': 0.20152544975280762, 'rewards/repeat_penalty/mean': 0.977634608745575, 'rewards/repeat_penalty/std': 0.009986236691474915, 'rewards/near_duplicate_penalty/mean': 0.977634608745575, 'rewards/near_duplicate_penalty/std': 0.009986236691474915, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.7809033989906311, 'rewards/total_composite/std': 0.1923426240682602, 'reward': 0.7809033989906311, 'reward_std': 0.1923426389694214, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0485701709985733, 'sampling/sampling_logp_difference/max': 1.2289419174194336, 'sampling/importance_sampling_ratio/min': 0.292602002620697, 'sampling/importance_sampling_ratio/mean': 1.0018235445022583, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.23541058972477913, 'clip_ratio/low_mean': 0.0035714285913854837, 'clip_ratio/low_min': 0.0035714285913854837, 'clip_ratio/high_mean': 0.04377159010618925, 'clip_ratio/high_max': 0.04377159010618925, 'clip_ratio/region_mean': 0.047343018697574735, 'reward_total_mean': 0.7809033989906311, 'reward_meter_mean': 0.9164797067642212, 'reward_meter_std': 0.021714089438319206, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.977634608745575, 'reward_repeat_penalty_std': 0.009986236691474915, 'reward_near_duplicate_penalty_mean': 0.977634608745575, 'reward_near_duplicate_penalty_std': 0.009986236691474915, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8487499952316284, 'reward_judge_quality_std': 0.20152544975280762, 'reward_total_composite_mean': 0.7809033989906311, 'reward_total_composite_std': 0.1923426240682602, 'epoch': 0.01}

 11%|█         | 347/3300 [1:16:57<8:41:28, 10.60s/it]INFO 04-12 12:03:09 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 348/3300 [1:17:10<9:26:29, 11.51s/it]2026-04-12 12:03:22,794 | INFO | train_grpo_train | metrics_logged mode=train step=348

                                                      
{'loss': -0.0082, 'grad_norm': 4.4145355224609375, 'learning_rate': 8.94848484848485e-06, 'num_tokens': 699408.0, 'completions/mean_length': 140.375, 'completions/min_length': 13.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 16.5, 'completions/min_terminated_length': 13.0, 'completions/max_terminated_length': 20.0, 'rewards/meter/mean': 0.48432457447052, 'rewards/meter/std': 0.4321945607662201, '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/lexical_plausibility/mean': 0.6881680488586426, 'rewards/lexical_plausibility/std': 0.3234315514564514, 'rewards/judge_quality/mean': 0.3500000238418579, 'rewards/judge_quality/std': 0.38172540068626404, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.23884817957878113, 'rewards/total_composite/std': 0.3289749026298523, 'reward': 0.23884817957878113, 'reward_std': 0.3289748728275299, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2495972216129303, 'sampling/sampling_logp_difference/max': 2.0360565185546875, 'sampling/importance_sampling_ratio/min': 0.13054250180721283, 'sampling/importance_sampling_ratio/mean': 0.9614812731742859, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8264362886548042, 'clip_ratio/low_mean': 0.07211538683623075, 'clip_ratio/low_min': 0.07211538683623075, 'clip_ratio/high_mean': 0.037593984976410866, 'clip_ratio/high_max': 0.037593984976410866, 'clip_ratio/region_mean': 0.10970937181264162, 'reward_total_mean': 0.23884817957878113, 'reward_meter_mean': 0.48432457447052, 'reward_meter_std': 0.4321945607662201, '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_lexical_plausibility_mean': 0.6881680488586426, 'reward_lexical_plausibility_std': 0.3234315514564514, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3500000238418579, 'reward_judge_quality_std': 0.38172540068626404, 'reward_total_composite_mean': 0.23884817957878113, 'reward_total_composite_std': 0.3289749026298523, 'epoch': 0.01}

 11%|█         | 348/3300 [1:17:10<9:26:29, 11.51s/it]INFO 04-12 12:03:23 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 349/3300 [1:17:20<9:07:31, 11.13s/it]2026-04-12 12:03:33,033 | INFO | train_grpo_train | metrics_logged mode=train step=349

                                                      
{'loss': 0.0046, 'grad_norm': 19.863420486450195, 'learning_rate': 8.945454545454546e-06, 'num_tokens': 700881.0, 'completions/mean_length': 40.125, 'completions/min_length': 35.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.125, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.8773584365844727, 'rewards/meter/std': 0.268604040145874, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.11649646610021591, 'rewards/repeat_penalty/mean': 0.9785360097885132, 'rewards/repeat_penalty/std': 0.02966911718249321, 'rewards/near_duplicate_penalty/mean': 0.9785360097885132, 'rewards/near_duplicate_penalty/std': 0.02966911718249321, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.30248257517814636, 'rewards/total_composite/std': 0.14327864348888397, 'reward': 0.30248257517814636, 'reward_std': 0.14327865839004517, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22356478869915009, 'sampling/sampling_logp_difference/max': 1.9981352090835571, 'sampling/importance_sampling_ratio/min': 0.1355878859758377, 'sampling/importance_sampling_ratio/mean': 1.017527461051941, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.713050052523613, 'clip_ratio/low_mean': 0.05018796864897013, 'clip_ratio/low_min': 0.05018796864897013, 'clip_ratio/high_mean': 0.15774736367166042, 'clip_ratio/high_max': 0.15774736367166042, 'clip_ratio/region_mean': 0.20793533232063055, 'reward_total_mean': 0.30248257517814636, 'reward_meter_mean': 0.8773584365844727, 'reward_meter_std': 0.268604040145874, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9785360097885132, 'reward_repeat_penalty_std': 0.02966911718249321, 'reward_near_duplicate_penalty_mean': 0.9785360097885132, 'reward_near_duplicate_penalty_std': 0.02966911718249321, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.11649646610021591, 'reward_total_composite_mean': 0.30248257517814636, 'reward_total_composite_std': 0.14327864348888397, 'epoch': 0.01}

 11%|█         | 349/3300 [1:17:21<9:07:31, 11.13s/it]INFO 04-12 12:03:33 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 350/3300 [1:17:31<8:55:54, 10.90s/it]2026-04-12 12:03:43,395 | INFO | train_grpo_train | metrics_logged mode=train step=350

                                                      
{'loss': 0.0112, 'grad_norm': 13.340145111083984, 'learning_rate': 8.942424242424243e-06, 'num_tokens': 702881.0, 'completions/mean_length': 63.0, 'completions/min_length': 56.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.0, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.47328752279281616, 'rewards/meter/std': 0.2505057752132416, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5587499737739563, 'rewards/judge_quality/std': 0.1950412094593048, 'rewards/repeat_penalty/mean': 0.9850226640701294, 'rewards/repeat_penalty/std': 0.026346338912844658, 'rewards/near_duplicate_penalty/mean': 0.9914755821228027, 'rewards/near_duplicate_penalty/std': 0.009483331814408302, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9933686852455139, 'rewards/distinct_2/std': 0.01875614933669567, 'rewards/total_composite/mean': 0.25895702838897705, 'rewards/total_composite/std': 0.1825934201478958, 'reward': 0.25895702838897705, 'reward_std': 0.18259340524673462, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1714884489774704, 'sampling/sampling_logp_difference/max': 2.6210365295410156, 'sampling/importance_sampling_ratio/min': 0.07272743433713913, 'sampling/importance_sampling_ratio/mean': 1.0101772546768188, 'sampling/importance_sampling_ratio/max': 1.8387874364852905, 'entropy': 1.3653811067342758, 'clip_ratio/low_mean': 0.10327029973268509, 'clip_ratio/low_min': 0.10327029973268509, 'clip_ratio/high_mean': 0.049912177957594395, 'clip_ratio/high_max': 0.049912177957594395, 'clip_ratio/region_mean': 0.15318247769027948, 'reward_total_mean': 0.25895702838897705, 'reward_meter_mean': 0.47328752279281616, 'reward_meter_std': 0.2505057752132416, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9850226640701294, 'reward_repeat_penalty_std': 0.026346338912844658, 'reward_near_duplicate_penalty_mean': 0.9914755821228027, 'reward_near_duplicate_penalty_std': 0.009483331814408302, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9933686852455139, 'reward_distinct_2_std': 0.01875614933669567, 'reward_judge_quality_mean': 0.5587499737739563, 'reward_judge_quality_std': 0.1950412094593048, 'reward_total_composite_mean': 0.25895702838897705, 'reward_total_composite_std': 0.1825934201478958, 'epoch': 0.01}

 11%|█         | 350/3300 [1:17:31<8:55:54, 10.90s/it]INFO 04-12 12:03:43 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


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

 15%|█▌        | 2/13 [00:10<00:59,  5.44s/it]

 23%|██▎       | 3/13 [00:18<01:01,  6.14s/it]

 31%|███       | 4/13 [00:29<01:14,  8.31s/it]

 38%|███▊      | 5/13 [00:42<01:17,  9.73s/it]

 46%|████▌     | 6/13 [00:53<01:11, 10.16s/it]

 54%|█████▍    | 7/13 [01:02<00:59,  9.86s/it]

 62%|██████▏   | 8/13 [01:15<00:53, 10.68s/it]

 69%|██████▉   | 9/13 [01:27<00:44, 11.10s/it]

 77%|███████▋  | 10/13 [01:40<00:35, 11.70s/it]

 85%|████████▍ | 11/13 [01:52<00:23, 11.79s/it]

 92%|█████████▏| 12/13 [02:04<00:11, 11.92s/it]

100%|██████████| 13/13 [02:14<00:00, 11.45s/it]2026-04-12 12:06:10,986 | INFO | train_grpo_train | metrics_logged mode=eval step=350

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 147.5508, 'eval_samples_per_second': 0.705, 'eval_steps_per_second': 0.088, 'eval_num_tokens': 702881.0, 'eval_completions/mean_length': 174.0, 'eval_completions/min_length': 37.76923076923077, 'eval_completions/max_length': 442.15384615384613, 'eval_completions/clipped_ratio': 0.08653846153846154, 'eval_completions/mean_terminated_length': 143.735353910006, 'eval_completions/min_terminated_length': 37.76923076923077, 'eval_completions/max_terminated_length': 317.84615384615387, 'eval_rewards/meter/mean': 0.5232180540378277, 'eval_rewards/meter/std': 0.36077991357216466, 'eval_rewards/count_adherence/mean': 0.8760176025904142, 'eval_rewards/count_adherence/std': 0.18586662010504648, 'eval_rewards/arabic_clean/mean': 0.7980769230769231, 'eval_rewards/arabic_clean/std': 0.3748629941390111, 'eval_rewards/lexical_plausibility/mean': 0.9659633957422696, 'eval_rewards/lexical_plausibility/std': 0.08558803060217403, 'eval_rewards/judge_quality/mean': 0.3753846058478722, 'eval_rewards/judge_quality/std': 0.19862179859326437, 'eval_rewards/repeat_penalty/mean': 0.805566246692951, 'eval_rewards/repeat_penalty/std': 0.28434896383147973, 'eval_rewards/near_duplicate_penalty/mean': 0.9476043581962585, 'eval_rewards/near_duplicate_penalty/std': 0.07918187882751226, 'eval_rewards/opening_diversity/mean': 0.9578615381167486, 'eval_rewards/opening_diversity/std': 0.08810311040053001, 'eval_rewards/distinct_2/mean': 0.8504216487591083, 'eval_rewards/distinct_2/std': 0.2567940211066833, 'eval_rewards/total_composite/mean': 0.1699994899905645, 'eval_rewards/total_composite/std': 0.17776415267815956, 'eval_reward': 0.1699994899905645, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.10092502030042502, 'eval_sampling/sampling_logp_difference/max': 1.2269824834970326, 'eval_sampling/importance_sampling_ratio/min': 0.2992986383346411, 'eval_sampling/importance_sampling_ratio/mean': 1.0261076780465932, 'eval_sampling/importance_sampling_ratio/max': 1.6199309734197764, 'eval_entropy': 1.4563191212140596, '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': 0.1699994899905645, 'eval_reward_meter_mean': 0.5232180540378277, 'eval_reward_meter_std': 0.36077991357216466, 'eval_reward_count_adherence_mean': 0.8760176025904142, 'eval_reward_count_adherence_std': 0.18586662010504648, 'eval_reward_arabic_clean_mean': 0.7980769230769231, 'eval_reward_arabic_clean_std': 0.3748629941390111, 'eval_reward_lexical_plausibility_mean': 0.9659633957422696, 'eval_reward_lexical_plausibility_std': 0.08558803060217403, 'eval_reward_repeat_penalty_mean': 0.805566246692951, 'eval_reward_repeat_penalty_std': 0.28434896383147973, 'eval_reward_near_duplicate_penalty_mean': 0.9476043581962585, 'eval_reward_near_duplicate_penalty_std': 0.07918187882751226, 'eval_reward_opening_diversity_mean': 0.9578615381167486, 'eval_reward_opening_diversity_std': 0.08810311040053001, 'eval_reward_distinct_2_mean': 0.8504216487591083, 'eval_reward_distinct_2_std': 0.2567940211066833, 'eval_reward_judge_quality_mean': 0.3753846058478722, 'eval_reward_judge_quality_std': 0.19862179859326437, 'eval_reward_total_composite_mean': 0.1699994899905645, 'eval_reward_total_composite_std': 0.17776415267815956, 'epoch': 0.01}

 11%|█         | 350/3300 [1:19:58<8:55:54, 10.90s/it]

100%|██████████| 13/13 [02:14<00:00, 11.45s/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(

 11%|█         | 351/3300 [1:20:11<45:37:38, 55.70s/it]2026-04-12 12:06:23,638 | INFO | train_grpo_train | metrics_logged mode=train step=351

                                                       
{'loss': 0.0526, 'grad_norm': 17.655532836914062, 'learning_rate': 8.93939393939394e-06, 'num_tokens': 704403.0, 'completions/mean_length': 37.25, 'completions/min_length': 35.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.25, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.7530195713043213, 'rewards/meter/std': 0.3564686179161072, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5424999594688416, 'rewards/judge_quality/std': 0.23705033957958221, 'rewards/repeat_penalty/mean': 0.952635645866394, 'rewards/repeat_penalty/std': 0.1334109902381897, 'rewards/near_duplicate_penalty/mean': 0.987712025642395, 'rewards/near_duplicate_penalty/std': 0.034202586859464645, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9898785352706909, 'rewards/distinct_2/std': 0.028627805411815643, 'rewards/total_composite/mean': 0.3259601593017578, 'rewards/total_composite/std': 0.2837991416454315, 'reward': 0.3259601593017578, 'reward_std': 0.2837991416454315, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1677749752998352, 'sampling/sampling_logp_difference/max': 1.7785744667053223, 'sampling/importance_sampling_ratio/min': 0.1688787192106247, 'sampling/importance_sampling_ratio/mean': 1.0243333578109741, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.269610472023487, 'clip_ratio/low_mean': 0.09796366095542908, 'clip_ratio/low_min': 0.09796366095542908, 'clip_ratio/high_mean': 0.10337301949039102, 'clip_ratio/high_max': 0.10337301949039102, 'clip_ratio/region_mean': 0.2013366804458201, 'reward_total_mean': 0.3259601593017578, 'reward_meter_mean': 0.7530195713043213, 'reward_meter_std': 0.3564686179161072, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.952635645866394, 'reward_repeat_penalty_std': 0.1334109902381897, 'reward_near_duplicate_penalty_mean': 0.987712025642395, 'reward_near_duplicate_penalty_std': 0.034202586859464645, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9898785352706909, 'reward_distinct_2_std': 0.028627805411815643, 'reward_judge_quality_mean': 0.5424999594688416, 'reward_judge_quality_std': 0.23705033957958221, 'reward_total_composite_mean': 0.3259601593017578, 'reward_total_composite_std': 0.2837991416454315, 'epoch': 0.01}

 11%|█         | 351/3300 [1:20:11<45:37:38, 55.70s/it]INFO 04-12 12:06:23 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 352/3300 [1:20:20<34:13:22, 41.79s/it]2026-04-12 12:06:32,971 | INFO | train_grpo_train | metrics_logged mode=train step=352

                                                       
{'loss': 0.0318, 'grad_norm': 22.35120391845703, 'learning_rate': 8.936363636363638e-06, 'num_tokens': 705882.0, 'completions/mean_length': 24.875, 'completions/min_length': 23.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.875, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.6298954486846924, 'rewards/meter/std': 0.44862452149391174, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.8594306707382202, 'rewards/lexical_plausibility/std': 0.2082936316728592, 'rewards/judge_quality/mean': 0.9237500429153442, 'rewards/judge_quality/std': 0.01060659158974886, 'rewards/repeat_penalty/mean': 0.7451828122138977, 'rewards/repeat_penalty/std': 0.013625085353851318, 'rewards/near_duplicate_penalty/mean': 0.9935770630836487, 'rewards/near_duplicate_penalty/std': 0.018166767433285713, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5832383036613464, 'rewards/total_composite/std': 0.4163356423377991, 'reward': 0.5832383036613464, 'reward_std': 0.4163356125354767, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15111209452152252, 'sampling/sampling_logp_difference/max': 1.638291835784912, 'sampling/importance_sampling_ratio/min': 0.19431167840957642, 'sampling/importance_sampling_ratio/mean': 0.9999074935913086, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7193644717335701, 'clip_ratio/low_mean': 0.04527243738994002, 'clip_ratio/low_min': 0.04527243738994002, 'clip_ratio/high_mean': 0.07821557903662324, 'clip_ratio/high_max': 0.07821557903662324, 'clip_ratio/region_mean': 0.12348801642656326, 'reward_total_mean': 0.5832383036613464, 'reward_meter_mean': 0.6298954486846924, 'reward_meter_std': 0.44862452149391174, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.8594306707382202, 'reward_lexical_plausibility_std': 0.2082936316728592, 'reward_repeat_penalty_mean': 0.7451828122138977, 'reward_repeat_penalty_std': 0.013625085353851318, 'reward_near_duplicate_penalty_mean': 0.9935770630836487, 'reward_near_duplicate_penalty_std': 0.018166767433285713, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.9237500429153442, 'reward_judge_quality_std': 0.01060659158974886, 'reward_total_composite_mean': 0.5832383036613464, 'reward_total_composite_std': 0.4163356423377991, 'epoch': 0.01}

 11%|█         | 352/3300 [1:20:20<34:13:22, 41.79s/it]INFO 04-12 12:06:33 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 353/3300 [1:20:29<26:10:32, 31.98s/it]2026-04-12 12:06:42,035 | INFO | train_grpo_train | metrics_logged mode=train step=353

                                                       
{'loss': 0.0458, 'grad_norm': 13.888103485107422, 'learning_rate': 8.933333333333333e-06, 'num_tokens': 707532.0, 'completions/mean_length': 45.25, 'completions/min_length': 39.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.25, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.9776734113693237, 'rewards/meter/std': 0.027861593291163445, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7712500095367432, 'rewards/judge_quality/std': 0.17771868407726288, 'rewards/repeat_penalty/mean': 0.9923610687255859, 'rewards/repeat_penalty/std': 0.014222200959920883, 'rewards/near_duplicate_penalty/mean': 0.9923610687255859, 'rewards/near_duplicate_penalty/std': 0.014222200959920883, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.7533683180809021, 'rewards/total_composite/std': 0.1725199967622757, 'reward': 0.7533683180809021, 'reward_std': 0.1725199818611145, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13646160066127777, 'sampling/sampling_logp_difference/max': 1.8458527326583862, 'sampling/importance_sampling_ratio/min': 0.15789063274860382, 'sampling/importance_sampling_ratio/mean': 1.0241644382476807, 'sampling/importance_sampling_ratio/max': 1.8805228471755981, 'entropy': 1.111404150724411, 'clip_ratio/low_mean': 0.054421061649918556, 'clip_ratio/low_min': 0.054421061649918556, 'clip_ratio/high_mean': 0.06493578106164932, 'clip_ratio/high_max': 0.06493578106164932, 'clip_ratio/region_mean': 0.11935684271156788, 'reward_total_mean': 0.7533683180809021, 'reward_meter_mean': 0.9776734113693237, 'reward_meter_std': 0.027861593291163445, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9923610687255859, 'reward_repeat_penalty_std': 0.014222200959920883, 'reward_near_duplicate_penalty_mean': 0.9923610687255859, 'reward_near_duplicate_penalty_std': 0.014222200959920883, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7712500095367432, 'reward_judge_quality_std': 0.17771868407726288, 'reward_total_composite_mean': 0.7533683180809021, 'reward_total_composite_std': 0.1725199967622757, 'epoch': 0.01}

 11%|█         | 353/3300 [1:20:30<26:10:32, 31.98s/it]INFO 04-12 12:06:42 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 354/3300 [1:20:40<20:48:54, 25.44s/it]2026-04-12 12:06:52,214 | INFO | train_grpo_train | metrics_logged mode=train step=354

                                                       
{'loss': 0.1136, 'grad_norm': 25.065343856811523, 'learning_rate': 8.930303030303032e-06, 'num_tokens': 709092.0, 'completions/mean_length': 32.0, 'completions/min_length': 27.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.0, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.8254019021987915, 'rewards/meter/std': 0.2102869600057602, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7574999928474426, 'rewards/judge_quality/std': 0.35025501251220703, 'rewards/repeat_penalty/mean': 0.9536963701248169, 'rewards/repeat_penalty/std': 0.059854354709386826, 'rewards/near_duplicate_penalty/mean': 0.9725993871688843, 'rewards/near_duplicate_penalty/std': 0.026196924969553947, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9797570705413818, 'rewards/distinct_2/std': 0.03748259320855141, 'rewards/total_composite/mean': 0.6245007514953613, 'rewards/total_composite/std': 0.3574490249156952, 'reward': 0.6245007514953613, 'reward_std': 0.3574490249156952, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16634072363376617, 'sampling/sampling_logp_difference/max': 2.130660057067871, 'sampling/importance_sampling_ratio/min': 0.11875887960195541, 'sampling/importance_sampling_ratio/mean': 0.9993553161621094, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9455316439270973, 'clip_ratio/low_mean': 0.06103830598294735, 'clip_ratio/low_min': 0.06103830598294735, 'clip_ratio/high_mean': 0.11087634321302176, 'clip_ratio/high_max': 0.11087634321302176, 'clip_ratio/region_mean': 0.1719146491959691, 'reward_total_mean': 0.6245007514953613, 'reward_meter_mean': 0.8254019021987915, 'reward_meter_std': 0.2102869600057602, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9536963701248169, 'reward_repeat_penalty_std': 0.059854354709386826, 'reward_near_duplicate_penalty_mean': 0.9725993871688843, 'reward_near_duplicate_penalty_std': 0.026196924969553947, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9797570705413818, 'reward_distinct_2_std': 0.03748259320855141, 'reward_judge_quality_mean': 0.7574999928474426, 'reward_judge_quality_std': 0.35025501251220703, 'reward_total_composite_mean': 0.6245007514953613, 'reward_total_composite_std': 0.3574490249156952, 'epoch': 0.01}

 11%|█         | 354/3300 [1:20:40<20:48:54, 25.44s/it]INFO 04-12 12:06:52 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 355/3300 [1:20:49<16:52:36, 20.63s/it]2026-04-12 12:07:01,634 | INFO | train_grpo_train | metrics_logged mode=train step=355

                                                       
{'loss': 0.0524, 'grad_norm': 16.550546646118164, 'learning_rate': 8.927272727272728e-06, 'num_tokens': 710833.0, 'completions/mean_length': 56.625, 'completions/min_length': 51.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.625, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.742302656173706, 'rewards/meter/std': 0.3499872088432312, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5625, 'rewards/judge_quality/std': 0.20858368277549744, 'rewards/repeat_penalty/mean': 0.9596279263496399, 'rewards/repeat_penalty/std': 0.032838139683008194, 'rewards/near_duplicate_penalty/mean': 0.9820369482040405, 'rewards/near_duplicate_penalty/std': 0.012260795570909977, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9771942496299744, 'rewards/distinct_2/std': 0.031511660665273666, 'rewards/total_composite/mean': 0.3991848826408386, 'rewards/total_composite/std': 0.23031742870807648, 'reward': 0.3991848826408386, 'reward_std': 0.23031742870807648, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17098252475261688, 'sampling/sampling_logp_difference/max': 1.5343976020812988, 'sampling/importance_sampling_ratio/min': 0.21558551490306854, 'sampling/importance_sampling_ratio/mean': 1.0290696620941162, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1870462223887444, 'clip_ratio/low_mean': 0.07203947380185127, 'clip_ratio/low_min': 0.07203947380185127, 'clip_ratio/high_mean': 0.1046310905367136, 'clip_ratio/high_max': 0.1046310905367136, 'clip_ratio/region_mean': 0.17667056433856487, 'reward_total_mean': 0.3991848826408386, 'reward_meter_mean': 0.742302656173706, 'reward_meter_std': 0.3499872088432312, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9596279263496399, 'reward_repeat_penalty_std': 0.032838139683008194, 'reward_near_duplicate_penalty_mean': 0.9820369482040405, 'reward_near_duplicate_penalty_std': 0.012260795570909977, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9771942496299744, 'reward_distinct_2_std': 0.031511660665273666, 'reward_judge_quality_mean': 0.5625, 'reward_judge_quality_std': 0.20858368277549744, 'reward_total_composite_mean': 0.3991848826408386, 'reward_total_composite_std': 0.23031742870807648, 'epoch': 0.01}

 11%|█         | 355/3300 [1:20:49<16:52:36, 20.63s/it]INFO 04-12 12:07:01 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 356/3300 [1:20:59<14:10:53, 17.34s/it]2026-04-12 12:07:11,297 | INFO | train_grpo_train | metrics_logged mode=train step=356

                                                       
{'loss': -0.0019, 'grad_norm': 11.859844207763672, 'learning_rate': 8.924242424242425e-06, 'num_tokens': 712869.0, 'completions/mean_length': 65.5, 'completions/min_length': 59.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.5, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.5503519177436829, 'rewards/meter/std': 0.3842824101448059, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5975000262260437, 'rewards/judge_quality/std': 0.35847896337509155, 'rewards/repeat_penalty/mean': 0.9002127051353455, 'rewards/repeat_penalty/std': 0.14802417159080505, 'rewards/near_duplicate_penalty/mean': 0.9710186123847961, 'rewards/near_duplicate_penalty/std': 0.027042049914598465, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9426940679550171, 'rewards/distinct_2/std': 0.08425553143024445, 'rewards/total_composite/mean': 0.2679823637008667, 'rewards/total_composite/std': 0.2850530445575714, 'reward': 0.2679823637008667, 'reward_std': 0.2850530445575714, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20014044642448425, 'sampling/sampling_logp_difference/max': 2.0784482955932617, 'sampling/importance_sampling_ratio/min': 0.12512421607971191, 'sampling/importance_sampling_ratio/mean': 1.0154919624328613, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6726450473070145, 'clip_ratio/low_mean': 0.14197781309485435, 'clip_ratio/low_min': 0.14197781309485435, 'clip_ratio/high_mean': 0.03549066372215748, 'clip_ratio/high_max': 0.03549066372215748, 'clip_ratio/region_mean': 0.17746847681701183, 'reward_total_mean': 0.2679823637008667, 'reward_meter_mean': 0.5503519177436829, 'reward_meter_std': 0.3842824101448059, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9002127051353455, 'reward_repeat_penalty_std': 0.14802417159080505, 'reward_near_duplicate_penalty_mean': 0.9710186123847961, 'reward_near_duplicate_penalty_std': 0.027042049914598465, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9426940679550171, 'reward_distinct_2_std': 0.08425553143024445, 'reward_judge_quality_mean': 0.5975000262260437, 'reward_judge_quality_std': 0.35847896337509155, 'reward_total_composite_mean': 0.2679823637008667, 'reward_total_composite_std': 0.2850530445575714, 'epoch': 0.01}

 11%|█         | 356/3300 [1:20:59<14:10:53, 17.34s/it]INFO 04-12 12:07:11 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 357/3300 [1:21:08<12:05:13, 14.79s/it]2026-04-12 12:07:20,118 | INFO | train_grpo_train | metrics_logged mode=train step=357

                                                       
{'loss': 0.0511, 'grad_norm': 16.52701759338379, 'learning_rate': 8.921212121212122e-06, 'num_tokens': 714417.0, 'completions/mean_length': 38.5, 'completions/min_length': 31.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.5, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.6863381266593933, 'rewards/meter/std': 0.37964722514152527, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4375, 'rewards/judge_quality/std': 0.1642080545425415, 'rewards/repeat_penalty/mean': 0.9445874691009521, 'rewards/repeat_penalty/std': 0.12602415680885315, 'rewards/near_duplicate_penalty/mean': 0.9799796342849731, 'rewards/near_duplicate_penalty/std': 0.032015178352594376, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.990384578704834, 'rewards/distinct_2/std': 0.027196412906050682, 'rewards/total_composite/mean': 0.2688857913017273, 'rewards/total_composite/std': 0.15353192389011383, 'reward': 0.2688857913017273, 'reward_std': 0.15353190898895264, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18544301390647888, 'sampling/sampling_logp_difference/max': 1.2146453857421875, 'sampling/importance_sampling_ratio/min': 0.29681524634361267, 'sampling/importance_sampling_ratio/mean': 1.0298609733581543, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.602674126625061, 'clip_ratio/low_mean': 0.060824490152299404, 'clip_ratio/low_min': 0.060824490152299404, 'clip_ratio/high_mean': 0.08745280560106039, 'clip_ratio/high_max': 0.08745280560106039, 'clip_ratio/region_mean': 0.1482772957533598, 'reward_total_mean': 0.2688857913017273, 'reward_meter_mean': 0.6863381266593933, 'reward_meter_std': 0.37964722514152527, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9445874691009521, 'reward_repeat_penalty_std': 0.12602415680885315, 'reward_near_duplicate_penalty_mean': 0.9799796342849731, 'reward_near_duplicate_penalty_std': 0.032015178352594376, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.990384578704834, 'reward_distinct_2_std': 0.027196412906050682, 'reward_judge_quality_mean': 0.4375, 'reward_judge_quality_std': 0.1642080545425415, 'reward_total_composite_mean': 0.2688857913017273, 'reward_total_composite_std': 0.15353192389011383, 'epoch': 0.01}

 11%|█         | 357/3300 [1:21:08<12:05:13, 14.79s/it]INFO 04-12 12:07:20 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 358/3300 [1:21:23<12:12:05, 14.93s/it]2026-04-12 12:07:35,389 | INFO | train_grpo_train | metrics_logged mode=train step=358

                                                       
{'loss': -0.1287, 'grad_norm': 2.8922390937805176, 'learning_rate': 8.91818181818182e-06, 'num_tokens': 717142.0, 'completions/mean_length': 287.625, 'completions/min_length': 179.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 212.83334350585938, 'completions/min_terminated_length': 179.0, 'completions/max_terminated_length': 254.0, 'rewards/meter/mean': 0.7647531032562256, 'rewards/meter/std': 0.3117305040359497, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.22519831359386444, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.8589385151863098, 'rewards/lexical_plausibility/std': 0.2416151463985443, 'rewards/judge_quality/mean': 0.17499999701976776, 'rewards/judge_quality/std': 0.11649647355079651, 'rewards/repeat_penalty/mean': 0.4069833755493164, 'rewards/repeat_penalty/std': 0.48276934027671814, 'rewards/near_duplicate_penalty/mean': 0.7356517314910889, 'rewards/near_duplicate_penalty/std': 0.2831598222255707, 'rewards/opening_diversity/mean': 0.576880693435669, 'rewards/opening_diversity/std': 0.38802570104599, 'rewards/distinct_2/mean': 0.4962750971317291, 'rewards/distinct_2/std': 0.4343658685684204, 'rewards/total_composite/mean': 0.0865609273314476, 'rewards/total_composite/std': 0.09177654981613159, 'reward': 0.0865609273314476, 'reward_std': 0.09177655726671219, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12771281599998474, 'sampling/sampling_logp_difference/max': 1.4083313941955566, 'sampling/importance_sampling_ratio/min': 0.27567747235298157, 'sampling/importance_sampling_ratio/mean': 1.0258456468582153, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1412059031426907, 'clip_ratio/low_mean': 0.01427165325731039, 'clip_ratio/low_min': 0.01427165325731039, 'clip_ratio/high_mean': 0.0731547859031707, 'clip_ratio/high_max': 0.0731547859031707, 'clip_ratio/region_mean': 0.0874264391604811, 'reward_total_mean': 0.0865609273314476, 'reward_meter_mean': 0.7647531032562256, 'reward_meter_std': 0.3117305040359497, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.22519831359386444, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.8589385151863098, 'reward_lexical_plausibility_std': 0.2416151463985443, 'reward_repeat_penalty_mean': 0.4069833755493164, 'reward_repeat_penalty_std': 0.48276934027671814, 'reward_near_duplicate_penalty_mean': 0.7356517314910889, 'reward_near_duplicate_penalty_std': 0.2831598222255707, 'reward_opening_diversity_mean': 0.576880693435669, 'reward_opening_diversity_std': 0.38802570104599, 'reward_distinct_2_mean': 0.4962750971317291, 'reward_distinct_2_std': 0.4343658685684204, 'reward_judge_quality_mean': 0.17499999701976776, 'reward_judge_quality_std': 0.11649647355079651, 'reward_total_composite_mean': 0.0865609273314476, 'reward_total_composite_std': 0.09177654981613159, 'epoch': 0.01}

 11%|█         | 358/3300 [1:21:23<12:12:05, 14.93s/it]INFO 04-12 12:07:35 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 359/3300 [1:21:32<10:46:08, 13.18s/it]2026-04-12 12:07:44,494 | INFO | train_grpo_train | metrics_logged mode=train step=359

                                                       
{'loss': 0.0992, 'grad_norm': 17.17350196838379, 'learning_rate': 8.915151515151515e-06, 'num_tokens': 718665.0, 'completions/mean_length': 43.375, 'completions/min_length': 36.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.375, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.7345317006111145, 'rewards/meter/std': 0.4362199902534485, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.71875, 'rewards/judge_quality/std': 0.30861154198646545, 'rewards/repeat_penalty/mean': 0.9529265761375427, 'rewards/repeat_penalty/std': 0.086003378033638, 'rewards/near_duplicate_penalty/mean': 0.9837132692337036, 'rewards/near_duplicate_penalty/std': 0.028064828366041183, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9673202633857727, 'rewards/distinct_2/std': 0.06421688199043274, 'rewards/total_composite/mean': 0.5463916063308716, 'rewards/total_composite/std': 0.43483734130859375, 'reward': 0.5463916063308716, 'reward_std': 0.43483734130859375, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15922391414642334, 'sampling/sampling_logp_difference/max': 1.474926471710205, 'sampling/importance_sampling_ratio/min': 0.22879555821418762, 'sampling/importance_sampling_ratio/mean': 0.9957095980644226, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0501817166805267, 'clip_ratio/low_mean': 0.05403903219848871, 'clip_ratio/low_min': 0.05403903219848871, 'clip_ratio/high_mean': 0.0999431386590004, 'clip_ratio/high_max': 0.0999431386590004, 'clip_ratio/region_mean': 0.1539821708574891, 'reward_total_mean': 0.5463916063308716, 'reward_meter_mean': 0.7345317006111145, 'reward_meter_std': 0.4362199902534485, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9529265761375427, 'reward_repeat_penalty_std': 0.086003378033638, 'reward_near_duplicate_penalty_mean': 0.9837132692337036, 'reward_near_duplicate_penalty_std': 0.028064828366041183, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9673202633857727, 'reward_distinct_2_std': 0.06421688199043274, 'reward_judge_quality_mean': 0.71875, 'reward_judge_quality_std': 0.30861154198646545, 'reward_total_composite_mean': 0.5463916063308716, 'reward_total_composite_std': 0.43483734130859375, 'epoch': 0.01}

 11%|█         | 359/3300 [1:21:32<10:46:08, 13.18s/it]INFO 04-12 12:07:44 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 360/3300 [1:21:40<9:32:26, 11.68s/it] 2026-04-12 12:07:52,682 | INFO | train_grpo_train | metrics_logged mode=train step=360

                                                      
{'loss': 0.1043, 'grad_norm': 28.60000991821289, 'learning_rate': 8.912121212121214e-06, 'num_tokens': 720341.0, 'completions/mean_length': 27.5, 'completions/min_length': 21.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.5, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.8417450785636902, 'rewards/meter/std': 0.28884682059288025, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.7730119824409485, 'rewards/lexical_plausibility/std': 0.2859584093093872, 'rewards/judge_quality/mean': 0.7725000381469727, 'rewards/judge_quality/std': 0.3027375340461731, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6335079669952393, 'rewards/total_composite/std': 0.3517366945743561, 'reward': 0.6335079669952393, 'reward_std': 0.3517366945743561, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1633785367012024, 'sampling/sampling_logp_difference/max': 1.6580560207366943, 'sampling/importance_sampling_ratio/min': 0.19050897657871246, 'sampling/importance_sampling_ratio/mean': 1.0053507089614868, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9615290611982346, 'clip_ratio/low_mean': 0.055898270569741726, 'clip_ratio/low_min': 0.055898270569741726, 'clip_ratio/high_mean': 0.11156580876559019, 'clip_ratio/high_max': 0.11156580876559019, 'clip_ratio/region_mean': 0.16746407933533192, 'reward_total_mean': 0.6335079669952393, 'reward_meter_mean': 0.8417450785636902, 'reward_meter_std': 0.28884682059288025, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.7730119824409485, 'reward_lexical_plausibility_std': 0.2859584093093872, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7725000381469727, 'reward_judge_quality_std': 0.3027375340461731, 'reward_total_composite_mean': 0.6335079669952393, 'reward_total_composite_std': 0.3517366945743561, 'epoch': 0.01}

 11%|█         | 360/3300 [1:21:40<9:32:26, 11.68s/it]INFO 04-12 12:07:52 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 361/3300 [1:21:51<9:15:16, 11.34s/it]2026-04-12 12:08:03,200 | INFO | train_grpo_train | metrics_logged mode=train step=361

                                                      
{'loss': 0.0668, 'grad_norm': 12.647197723388672, 'learning_rate': 8.90909090909091e-06, 'num_tokens': 722153.0, 'completions/mean_length': 63.5, 'completions/min_length': 58.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.5, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.6334871053695679, 'rewards/meter/std': 0.3794703781604767, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.34999996423721313, 'rewards/judge_quality/std': 0.13093073666095734, 'rewards/repeat_penalty/mean': 0.9894030690193176, 'rewards/repeat_penalty/std': 0.01291202288120985, 'rewards/near_duplicate_penalty/mean': 0.9894030690193176, 'rewards/near_duplicate_penalty/std': 0.01291202288120985, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2501254081726074, 'rewards/total_composite/std': 0.18298527598381042, 'reward': 0.2501254081726074, 'reward_std': 0.18298527598381042, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19529755413532257, 'sampling/sampling_logp_difference/max': 1.879929542541504, 'sampling/importance_sampling_ratio/min': 0.15260085463523865, 'sampling/importance_sampling_ratio/mean': 1.057754397392273, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.251577615737915, 'clip_ratio/low_mean': 0.11461598798632622, 'clip_ratio/low_min': 0.11461598798632622, 'clip_ratio/high_mean': 0.07472979836165905, 'clip_ratio/high_max': 0.07472979836165905, 'clip_ratio/region_mean': 0.18934578634798527, 'reward_total_mean': 0.2501254081726074, 'reward_meter_mean': 0.6334871053695679, 'reward_meter_std': 0.3794703781604767, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.9894030690193176, 'reward_repeat_penalty_std': 0.01291202288120985, 'reward_near_duplicate_penalty_mean': 0.9894030690193176, 'reward_near_duplicate_penalty_std': 0.01291202288120985, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.34999996423721313, 'reward_judge_quality_std': 0.13093073666095734, 'reward_total_composite_mean': 0.2501254081726074, 'reward_total_composite_std': 0.18298527598381042, 'epoch': 0.01}

 11%|█         | 361/3300 [1:21:51<9:15:16, 11.34s/it]INFO 04-12 12:08:03 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 362/3300 [1:21:59<8:38:14, 10.58s/it]2026-04-12 12:08:12,031 | INFO | train_grpo_train | metrics_logged mode=train step=362

                                                      
{'loss': 0.079, 'grad_norm': 12.049482345581055, 'learning_rate': 8.906060606060607e-06, 'num_tokens': 723753.0, 'completions/mean_length': 41.0, 'completions/min_length': 37.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.0, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.9270881414413452, 'rewards/meter/std': 0.05334731563925743, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6225000023841858, 'rewards/judge_quality/std': 0.3285357356071472, 'rewards/repeat_penalty/mean': 0.854789137840271, 'rewards/repeat_penalty/std': 0.16840942203998566, 'rewards/near_duplicate_penalty/mean': 0.9022011160850525, 'rewards/near_duplicate_penalty/std': 0.09423723071813583, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.966063380241394, 'rewards/distinct_2/std': 0.04683702811598778, 'rewards/total_composite/mean': 0.5747998356819153, 'rewards/total_composite/std': 0.29921311140060425, 'reward': 0.5747998356819153, 'reward_std': 0.29921311140060425, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15688970685005188, 'sampling/sampling_logp_difference/max': 2.671311855316162, 'sampling/importance_sampling_ratio/min': 0.06916143745183945, 'sampling/importance_sampling_ratio/mean': 1.0157854557037354, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.014362320303917, 'clip_ratio/low_mean': 0.06113460287451744, 'clip_ratio/low_min': 0.06113460287451744, 'clip_ratio/high_mean': 0.10501393862068653, 'clip_ratio/high_max': 0.10501393862068653, 'clip_ratio/region_mean': 0.16614854149520397, 'reward_total_mean': 0.5747998356819153, 'reward_meter_mean': 0.9270881414413452, 'reward_meter_std': 0.05334731563925743, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.854789137840271, 'reward_repeat_penalty_std': 0.16840942203998566, 'reward_near_duplicate_penalty_mean': 0.9022011160850525, 'reward_near_duplicate_penalty_std': 0.09423723071813583, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.966063380241394, 'reward_distinct_2_std': 0.04683702811598778, 'reward_judge_quality_mean': 0.6225000023841858, 'reward_judge_quality_std': 0.3285357356071472, 'reward_total_composite_mean': 0.5747998356819153, 'reward_total_composite_std': 0.29921311140060425, 'epoch': 0.01}

 11%|█         | 362/3300 [1:22:00<8:38:14, 10.58s/it]INFO 04-12 12:08:12 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 363/3300 [1:22:15<9:49:07, 12.04s/it]2026-04-12 12:08:27,456 | INFO | train_grpo_train | metrics_logged mode=train step=363

                                                      
{'loss': -0.1432, 'grad_norm': 3.6705923080444336, 'learning_rate': 8.903030303030304e-06, 'num_tokens': 725791.0, 'completions/mean_length': 202.75, 'completions/min_length': 82.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 99.66667175292969, 'completions/min_terminated_length': 82.0, 'completions/max_terminated_length': 118.0, 'rewards/meter/mean': 0.45061933994293213, 'rewards/meter/std': 0.3779248297214508, 'rewards/count_adherence/mean': 0.7999999523162842, 'rewards/count_adherence/std': 0.33806172013282776, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9334577322006226, 'rewards/lexical_plausibility/std': 0.14861370623111725, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.1752549111843109, 'rewards/repeat_penalty/mean': 0.9784387350082397, 'rewards/repeat_penalty/std': 0.020184149965643883, 'rewards/near_duplicate_penalty/mean': 0.9864121675491333, 'rewards/near_duplicate_penalty/std': 0.012301817536354065, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.991870105266571, 'rewards/distinct_2/std': 0.012649363838136196, 'rewards/total_composite/mean': 0.164620041847229, 'rewards/total_composite/std': 0.158761665225029, 'reward': 0.164620041847229, 'reward_std': 0.1587616503238678, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1910051852464676, 'sampling/sampling_logp_difference/max': 1.5631217956542969, 'sampling/importance_sampling_ratio/min': 0.20948109030723572, 'sampling/importance_sampling_ratio/mean': 1.030908465385437, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5991112738847733, 'clip_ratio/low_mean': 0.027108434587717056, 'clip_ratio/low_min': 0.027108434587717056, 'clip_ratio/high_mean': 0.1284009926021099, 'clip_ratio/high_max': 0.1284009926021099, 'clip_ratio/region_mean': 0.15550942718982697, 'reward_total_mean': 0.164620041847229, 'reward_meter_mean': 0.45061933994293213, 'reward_meter_std': 0.3779248297214508, 'reward_count_adherence_mean': 0.7999999523162842, 'reward_count_adherence_std': 0.33806172013282776, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9334577322006226, 'reward_lexical_plausibility_std': 0.14861370623111725, 'reward_repeat_penalty_mean': 0.9784387350082397, 'reward_repeat_penalty_std': 0.020184149965643883, 'reward_near_duplicate_penalty_mean': 0.9864121675491333, 'reward_near_duplicate_penalty_std': 0.012301817536354065, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.991870105266571, 'reward_distinct_2_std': 0.012649363838136196, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.1752549111843109, 'reward_total_composite_mean': 0.164620041847229, 'reward_total_composite_std': 0.158761665225029, 'epoch': 0.01}

 11%|█         | 363/3300 [1:22:15<9:49:07, 12.04s/it]INFO 04-12 12:08:27 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 364/3300 [1:22:26<9:28:10, 11.61s/it]2026-04-12 12:08:38,074 | INFO | train_grpo_train | metrics_logged mode=train step=364

                                                      
{'loss': 0.0242, 'grad_norm': 15.96010971069336, 'learning_rate': 8.900000000000001e-06, 'num_tokens': 727337.0, 'completions/mean_length': 36.25, 'completions/min_length': 31.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.25, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.5769586563110352, 'rewards/meter/std': 0.3881268799304962, '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/lexical_plausibility/mean': 0.9822304248809814, 'rewards/lexical_plausibility/std': 0.0502600334584713, 'rewards/judge_quality/mean': 0.4124999940395355, 'rewards/judge_quality/std': 0.1685018092393875, 'rewards/repeat_penalty/mean': 0.9903969168663025, 'rewards/repeat_penalty/std': 0.015086086466908455, 'rewards/near_duplicate_penalty/mean': 0.9903969168663025, 'rewards/near_duplicate_penalty/std': 0.015086086466908455, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.236695796251297, 'rewards/total_composite/std': 0.24888162314891815, 'reward': 0.236695796251297, 'reward_std': 0.24888163805007935, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2014869749546051, 'sampling/sampling_logp_difference/max': 1.3099784851074219, 'sampling/importance_sampling_ratio/min': 0.26982587575912476, 'sampling/importance_sampling_ratio/mean': 1.034716010093689, 'sampling/importance_sampling_ratio/max': 1.9667059183120728, 'entropy': 1.6048991158604622, 'clip_ratio/low_mean': 0.07555138133466244, 'clip_ratio/low_min': 0.07555138133466244, 'clip_ratio/high_mean': 0.06635742541402578, 'clip_ratio/high_max': 0.06635742541402578, 'clip_ratio/region_mean': 0.14190880674868822, 'reward_total_mean': 0.236695796251297, 'reward_meter_mean': 0.5769586563110352, 'reward_meter_std': 0.3881268799304962, '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_lexical_plausibility_mean': 0.9822304248809814, 'reward_lexical_plausibility_std': 0.0502600334584713, 'reward_repeat_penalty_mean': 0.9903969168663025, 'reward_repeat_penalty_std': 0.015086086466908455, 'reward_near_duplicate_penalty_mean': 0.9903969168663025, 'reward_near_duplicate_penalty_std': 0.015086086466908455, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4124999940395355, 'reward_judge_quality_std': 0.1685018092393875, 'reward_total_composite_mean': 0.236695796251297, 'reward_total_composite_std': 0.24888162314891815, 'epoch': 0.01}

 11%|█         | 364/3300 [1:22:26<9:28:10, 11.61s/it]INFO 04-12 12:08:38 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 365/3300 [1:22:34<8:47:55, 10.79s/it]2026-04-12 12:08:46,960 | INFO | train_grpo_train | metrics_logged mode=train step=365

                                                      
{'loss': 0.0671, 'grad_norm': 17.79912757873535, 'learning_rate': 8.896969696969697e-06, 'num_tokens': 728700.0, 'completions/mean_length': 36.375, 'completions/min_length': 33.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.375, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.9167230725288391, 'rewards/meter/std': 0.14205153286457062, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5537500381469727, 'rewards/judge_quality/std': 0.18469570577144623, 'rewards/repeat_penalty/mean': 0.8419625759124756, 'rewards/repeat_penalty/std': 0.2098531872034073, 'rewards/near_duplicate_penalty/mean': 0.9634506702423096, 'rewards/near_duplicate_penalty/std': 0.04839060828089714, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9482118487358093, 'rewards/distinct_2/std': 0.09005054086446762, 'rewards/total_composite/mean': 0.5047616958618164, 'rewards/total_composite/std': 0.19028246402740479, 'reward': 0.5047616958618164, 'reward_std': 0.19028246402740479, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1717405915260315, 'sampling/sampling_logp_difference/max': 2.181375503540039, 'sampling/importance_sampling_ratio/min': 0.11288614571094513, 'sampling/importance_sampling_ratio/mean': 1.0314292907714844, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.60911326110363, 'clip_ratio/low_mean': 0.10354327782988548, 'clip_ratio/low_min': 0.10354327782988548, 'clip_ratio/high_mean': 0.054306723177433014, 'clip_ratio/high_max': 0.054306723177433014, 'clip_ratio/region_mean': 0.1578500010073185, 'reward_total_mean': 0.5047616958618164, 'reward_meter_mean': 0.9167230725288391, 'reward_meter_std': 0.14205153286457062, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8419625759124756, 'reward_repeat_penalty_std': 0.2098531872034073, 'reward_near_duplicate_penalty_mean': 0.9634506702423096, 'reward_near_duplicate_penalty_std': 0.04839060828089714, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9482118487358093, 'reward_distinct_2_std': 0.09005054086446762, 'reward_judge_quality_mean': 0.5537500381469727, 'reward_judge_quality_std': 0.18469570577144623, 'reward_total_composite_mean': 0.5047616958618164, 'reward_total_composite_std': 0.19028246402740479, 'epoch': 0.01}

 11%|█         | 365/3300 [1:22:34<8:47:55, 10.79s/it]INFO 04-12 12:08:47 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 366/3300 [1:22:48<9:23:29, 11.52s/it]2026-04-12 12:09:00,186 | INFO | train_grpo_train | metrics_logged mode=train step=366

                                                      
{'loss': -0.0659, 'grad_norm': 2.6969549655914307, 'learning_rate': 8.893939393939394e-06, 'num_tokens': 730246.0, 'completions/mean_length': 164.25, 'completions/min_length': 38.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 48.333335876464844, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.6664940118789673, 'rewards/meter/std': 0.4521222710609436, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.8626080751419067, 'rewards/lexical_plausibility/std': 0.2585444450378418, 'rewards/judge_quality/mean': 0.5262500047683716, 'rewards/judge_quality/std': 0.38101136684417725, 'rewards/repeat_penalty/mean': 0.9925903081893921, 'rewards/repeat_penalty/std': 0.01526072807610035, 'rewards/near_duplicate_penalty/mean': 0.9925903081893921, 'rewards/near_duplicate_penalty/std': 0.01526072807610035, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.28297725319862366, 'rewards/total_composite/std': 0.35254907608032227, 'reward': 0.28297725319862366, 'reward_std': 0.3525490462779999, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20700275897979736, 'sampling/sampling_logp_difference/max': 1.6284193992614746, 'sampling/importance_sampling_ratio/min': 0.19623950123786926, 'sampling/importance_sampling_ratio/mean': 0.998453676700592, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2886315882205963, 'clip_ratio/low_mean': 0.06262258999049664, 'clip_ratio/low_min': 0.06262258999049664, 'clip_ratio/high_mean': 0.07659623958170414, 'clip_ratio/high_max': 0.07659623958170414, 'clip_ratio/region_mean': 0.13921882957220078, 'reward_total_mean': 0.28297725319862366, 'reward_meter_mean': 0.6664940118789673, 'reward_meter_std': 0.4521222710609436, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.8626080751419067, 'reward_lexical_plausibility_std': 0.2585444450378418, 'reward_repeat_penalty_mean': 0.9925903081893921, 'reward_repeat_penalty_std': 0.01526072807610035, 'reward_near_duplicate_penalty_mean': 0.9925903081893921, 'reward_near_duplicate_penalty_std': 0.01526072807610035, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5262500047683716, 'reward_judge_quality_std': 0.38101136684417725, 'reward_total_composite_mean': 0.28297725319862366, 'reward_total_composite_std': 0.35254907608032227, 'epoch': 0.01}

 11%|█         | 366/3300 [1:22:48<9:23:29, 11.52s/it]INFO 04-12 12:09:00 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 367/3300 [1:23:02<10:08:21, 12.45s/it]2026-04-12 12:09:14,782 | INFO | train_grpo_train | metrics_logged mode=train step=367

                                                       
{'loss': -0.025, 'grad_norm': 6.268082141876221, 'learning_rate': 8.890909090909091e-06, 'num_tokens': 731992.0, 'completions/mean_length': 117.25, 'completions/min_length': 54.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 60.857147216796875, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.356907457113266, 'rewards/meter/std': 0.3961610794067383, '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/lexical_plausibility/mean': 0.9265891909599304, 'rewards/lexical_plausibility/std': 0.20763714611530304, 'rewards/judge_quality/mean': 0.48375001549720764, 'rewards/judge_quality/std': 0.3717886507511139, 'rewards/repeat_penalty/mean': 0.9814070463180542, 'rewards/repeat_penalty/std': 0.012411732226610184, 'rewards/near_duplicate_penalty/mean': 0.9814070463180542, 'rewards/near_duplicate_penalty/std': 0.012411732226610184, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.27941232919692993, 'rewards/total_composite/std': 0.354913592338562, 'reward': 0.27941232919692993, 'reward_std': 0.354913592338562, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20100927352905273, 'sampling/sampling_logp_difference/max': 1.6287550926208496, 'sampling/importance_sampling_ratio/min': 0.19617362320423126, 'sampling/importance_sampling_ratio/mean': 1.0245027542114258, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6056828796863556, 'clip_ratio/low_mean': 0.09891764167696238, 'clip_ratio/low_min': 0.09891764167696238, 'clip_ratio/high_mean': 0.0677736010402441, 'clip_ratio/high_max': 0.0677736010402441, 'clip_ratio/region_mean': 0.16669124271720648, 'reward_total_mean': 0.27941232919692993, 'reward_meter_mean': 0.356907457113266, 'reward_meter_std': 0.3961610794067383, '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_lexical_plausibility_mean': 0.9265891909599304, 'reward_lexical_plausibility_std': 0.20763714611530304, 'reward_repeat_penalty_mean': 0.9814070463180542, 'reward_repeat_penalty_std': 0.012411732226610184, 'reward_near_duplicate_penalty_mean': 0.9814070463180542, 'reward_near_duplicate_penalty_std': 0.012411732226610184, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.48375001549720764, 'reward_judge_quality_std': 0.3717886507511139, 'reward_total_composite_mean': 0.27941232919692993, 'reward_total_composite_std': 0.354913592338562, 'epoch': 0.01}

 11%|█         | 367/3300 [1:23:02<10:08:21, 12.45s/it]INFO 04-12 12:09:15 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 368/3300 [1:23:11<9:13:08, 11.32s/it] 2026-04-12 12:09:23,475 | INFO | train_grpo_train | metrics_logged mode=train step=368

                                                      
{'loss': 0.1164, 'grad_norm': 21.416719436645508, 'learning_rate': 8.887878787878789e-06, 'num_tokens': 733654.0, 'completions/mean_length': 38.75, 'completions/min_length': 34.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.75, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.7224254012107849, 'rewards/meter/std': 0.34737133979797363, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.41249996423721313, 'rewards/judge_quality/std': 0.14078859984874725, 'rewards/repeat_penalty/mean': 0.9940342903137207, 'rewards/repeat_penalty/std': 0.008988606743514538, 'rewards/near_duplicate_penalty/mean': 0.9940342903137207, 'rewards/near_duplicate_penalty/std': 0.008988606743514538, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2602403461933136, 'rewards/total_composite/std': 0.18446087837219238, 'reward': 0.2602403461933136, 'reward_std': 0.18446087837219238, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18854431807994843, 'sampling/sampling_logp_difference/max': 1.5036060810089111, 'sampling/importance_sampling_ratio/min': 0.22232699394226074, 'sampling/importance_sampling_ratio/mean': 0.9972913265228271, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5983603745698929, 'clip_ratio/low_mean': 0.04927333444356918, 'clip_ratio/low_min': 0.04927333444356918, 'clip_ratio/high_mean': 0.11093856766819954, 'clip_ratio/high_max': 0.11093856766819954, 'clip_ratio/region_mean': 0.16021190211176872, 'reward_total_mean': 0.2602403461933136, 'reward_meter_mean': 0.7224254012107849, 'reward_meter_std': 0.34737133979797363, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9940342903137207, 'reward_repeat_penalty_std': 0.008988606743514538, 'reward_near_duplicate_penalty_mean': 0.9940342903137207, 'reward_near_duplicate_penalty_std': 0.008988606743514538, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.41249996423721313, 'reward_judge_quality_std': 0.14078859984874725, 'reward_total_composite_mean': 0.2602403461933136, 'reward_total_composite_std': 0.18446087837219238, 'epoch': 0.01}

 11%|█         | 368/3300 [1:23:11<9:13:08, 11.32s/it]INFO 04-12 12:09:23 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 369/3300 [1:23:21<8:49:28, 10.84s/it]2026-04-12 12:09:33,190 | INFO | train_grpo_train | metrics_logged mode=train step=369

                                                      
{'loss': 0.003, 'grad_norm': 11.890995025634766, 'learning_rate': 8.884848484848486e-06, 'num_tokens': 735556.0, 'completions/mean_length': 66.75, 'completions/min_length': 57.0, 'completions/max_length': 84.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.75, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.7550082206726074, 'rewards/meter/std': 0.34770846366882324, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9890648722648621, 'rewards/lexical_plausibility/std': 0.03092925436794758, 'rewards/judge_quality/mean': 0.4375, 'rewards/judge_quality/std': 0.0353553369641304, 'rewards/repeat_penalty/mean': 0.887050986289978, 'rewards/repeat_penalty/std': 0.21987536549568176, 'rewards/near_duplicate_penalty/mean': 0.9636750221252441, 'rewards/near_duplicate_penalty/std': 0.04604289308190346, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9254437685012817, 'rewards/distinct_2/std': 0.16680213809013367, 'rewards/total_composite/mean': 0.32750314474105835, 'rewards/total_composite/std': 0.15115393698215485, 'reward': 0.32750314474105835, 'reward_std': 0.15115393698215485, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14748868346214294, 'sampling/sampling_logp_difference/max': 1.5324044227600098, 'sampling/importance_sampling_ratio/min': 0.21601565182209015, 'sampling/importance_sampling_ratio/mean': 1.0249985456466675, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.081573411822319, 'clip_ratio/low_mean': 0.029146634973585606, 'clip_ratio/low_min': 0.029146634973585606, 'clip_ratio/high_mean': 0.0987002206966281, 'clip_ratio/high_max': 0.0987002206966281, 'clip_ratio/region_mean': 0.1278468556702137, 'reward_total_mean': 0.32750314474105835, 'reward_meter_mean': 0.7550082206726074, 'reward_meter_std': 0.34770846366882324, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9890648722648621, 'reward_lexical_plausibility_std': 0.03092925436794758, 'reward_repeat_penalty_mean': 0.887050986289978, 'reward_repeat_penalty_std': 0.21987536549568176, 'reward_near_duplicate_penalty_mean': 0.9636750221252441, 'reward_near_duplicate_penalty_std': 0.04604289308190346, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9254437685012817, 'reward_distinct_2_std': 0.16680213809013367, 'reward_judge_quality_mean': 0.4375, 'reward_judge_quality_std': 0.0353553369641304, 'reward_total_composite_mean': 0.32750314474105835, 'reward_total_composite_std': 0.15115393698215485, 'epoch': 0.01}

 11%|█         | 369/3300 [1:23:21<8:49:28, 10.84s/it]INFO 04-12 12:09:33 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 370/3300 [1:23:34<9:30:57, 11.69s/it]2026-04-12 12:09:46,885 | INFO | train_grpo_train | metrics_logged mode=train step=370

                                                      
{'loss': -0.073, 'grad_norm': 3.1572275161743164, 'learning_rate': 8.881818181818183e-06, 'num_tokens': 737047.0, 'completions/mean_length': 158.375, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 40.5, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.8391140699386597, 'rewards/meter/std': 0.3399761915206909, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.7502424716949463, 'rewards/lexical_plausibility/std': 0.26979056000709534, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.18077215552330017, 'rewards/repeat_penalty/mean': 0.7485777735710144, 'rewards/repeat_penalty/std': 0.36676982045173645, 'rewards/near_duplicate_penalty/mean': 0.9593077301979065, 'rewards/near_duplicate_penalty/std': 0.04626181349158287, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.3471825420856476, 'rewards/distinct_2/mean': 0.9055330753326416, 'rewards/distinct_2/std': 0.1304432600736618, 'rewards/total_composite/mean': 0.23798933625221252, 'rewards/total_composite/std': 0.18960289657115936, 'reward': 0.23798933625221252, 'reward_std': 0.18960291147232056, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2261786013841629, 'sampling/sampling_logp_difference/max': 1.7096266746520996, 'sampling/importance_sampling_ratio/min': 0.18093332648277283, 'sampling/importance_sampling_ratio/mean': 1.053892970085144, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4391814470291138, 'clip_ratio/low_mean': 0.04377736151218414, 'clip_ratio/low_min': 0.04377736151218414, 'clip_ratio/high_mean': 0.06570638995617628, 'clip_ratio/high_max': 0.06570638995617628, 'clip_ratio/region_mean': 0.10948375146836042, 'reward_total_mean': 0.23798933625221252, 'reward_meter_mean': 0.8391140699386597, 'reward_meter_std': 0.3399761915206909, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.7502424716949463, 'reward_lexical_plausibility_std': 0.26979056000709534, 'reward_repeat_penalty_mean': 0.7485777735710144, 'reward_repeat_penalty_std': 0.36676982045173645, 'reward_near_duplicate_penalty_mean': 0.9593077301979065, 'reward_near_duplicate_penalty_std': 0.04626181349158287, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.3471825420856476, 'reward_distinct_2_mean': 0.9055330753326416, 'reward_distinct_2_std': 0.1304432600736618, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.18077215552330017, 'reward_total_composite_mean': 0.23798933625221252, 'reward_total_composite_std': 0.18960289657115936, 'epoch': 0.01}

 11%|█         | 370/3300 [1:23:34<9:30:57, 11.69s/it]INFO 04-12 12:09:47 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 371/3300 [1:23:49<10:11:44, 12.53s/it]2026-04-12 12:10:01,363 | INFO | train_grpo_train | metrics_logged mode=train step=371

                                                       
{'loss': -0.1166, 'grad_norm': 2.637321710586548, 'learning_rate': 8.87878787878788e-06, 'num_tokens': 740804.0, 'completions/mean_length': 311.625, 'completions/min_length': 232.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 283.0, 'completions/min_terminated_length': 232.0, 'completions/max_terminated_length': 337.0, 'rewards/meter/mean': 0.7757177948951721, 'rewards/meter/std': 0.344845712184906, 'rewards/count_adherence/mean': 0.7604166269302368, 'rewards/count_adherence/std': 0.05340581759810448, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9530562162399292, 'rewards/lexical_plausibility/std': 0.1327771246433258, 'rewards/judge_quality/mean': 0.21250000596046448, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.22648166120052338, 'rewards/repeat_penalty/std': 0.34410199522972107, 'rewards/near_duplicate_penalty/mean': 0.7649456262588501, 'rewards/near_duplicate_penalty/std': 0.14382021129131317, 'rewards/opening_diversity/mean': 0.8883928656578064, 'rewards/opening_diversity/std': 0.1704024374485016, 'rewards/distinct_2/mean': 0.2804200053215027, 'rewards/distinct_2/std': 0.3543919622898102, 'rewards/total_composite/mean': 0.13439007103443146, 'rewards/total_composite/std': 0.08367897570133209, 'reward': 0.13439007103443146, 'reward_std': 0.08367897570133209, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0919528529047966, 'sampling/sampling_logp_difference/max': 2.24222469329834, 'sampling/importance_sampling_ratio/min': 0.10622192919254303, 'sampling/importance_sampling_ratio/mean': 1.019260287284851, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6797917410731316, 'clip_ratio/low_mean': 0.02716411859728396, 'clip_ratio/low_min': 0.02716411859728396, 'clip_ratio/high_mean': 0.04265493992716074, 'clip_ratio/high_max': 0.04265493992716074, 'clip_ratio/region_mean': 0.0698190585244447, 'reward_total_mean': 0.13439007103443146, 'reward_meter_mean': 0.7757177948951721, 'reward_meter_std': 0.344845712184906, 'reward_count_adherence_mean': 0.7604166269302368, 'reward_count_adherence_std': 0.05340581759810448, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9530562162399292, 'reward_lexical_plausibility_std': 0.1327771246433258, 'reward_repeat_penalty_mean': 0.22648166120052338, 'reward_repeat_penalty_std': 0.34410199522972107, 'reward_near_duplicate_penalty_mean': 0.7649456262588501, 'reward_near_duplicate_penalty_std': 0.14382021129131317, 'reward_opening_diversity_mean': 0.8883928656578064, 'reward_opening_diversity_std': 0.1704024374485016, 'reward_distinct_2_mean': 0.2804200053215027, 'reward_distinct_2_std': 0.3543919622898102, 'reward_judge_quality_mean': 0.21250000596046448, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.13439007103443146, 'reward_total_composite_std': 0.08367897570133209, 'epoch': 0.01}

 11%|█         | 371/3300 [1:23:49<10:11:44, 12.53s/it]INFO 04-12 12:10:01 [block_pool.py:378] Successfully reset prefix cache

 11%|█▏        | 372/3300 [1:24:00<9:54:48, 12.19s/it] 2026-04-12 12:10:12,754 | INFO | train_grpo_train | metrics_logged mode=train step=372

                                                      
{'loss': 0.0657, 'grad_norm': 7.444216728210449, 'learning_rate': 8.875757575757576e-06, 'num_tokens': 743324.0, 'completions/mean_length': 111.0, 'completions/min_length': 103.0, 'completions/max_length': 137.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 111.0, 'completions/min_terminated_length': 103.0, 'completions/max_terminated_length': 137.0, 'rewards/meter/mean': 0.7907493114471436, 'rewards/meter/std': 0.28211304545402527, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.35874998569488525, 'rewards/judge_quality/std': 0.2547512948513031, 'rewards/repeat_penalty/mean': 0.7878605723381042, 'rewards/repeat_penalty/std': 0.3069226145744324, 'rewards/near_duplicate_penalty/mean': 0.9179651141166687, 'rewards/near_duplicate_penalty/std': 0.13240814208984375, 'rewards/opening_diversity/mean': 0.8999999761581421, 'rewards/opening_diversity/std': 0.24494896829128265, 'rewards/distinct_2/mean': 0.8616527318954468, 'rewards/distinct_2/std': 0.2173728346824646, 'rewards/total_composite/mean': 0.29524925351142883, 'rewards/total_composite/std': 0.27520182728767395, 'reward': 0.29524925351142883, 'reward_std': 0.27520182728767395, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16958843171596527, 'sampling/sampling_logp_difference/max': 1.742572546005249, 'sampling/importance_sampling_ratio/min': 0.17506945133209229, 'sampling/importance_sampling_ratio/mean': 1.0220175981521606, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2887337058782578, 'clip_ratio/low_mean': 0.06503461673855782, 'clip_ratio/low_min': 0.06503461673855782, 'clip_ratio/high_mean': 0.07919359020888805, 'clip_ratio/high_max': 0.07919359020888805, 'clip_ratio/region_mean': 0.14422820694744587, 'reward_total_mean': 0.29524925351142883, 'reward_meter_mean': 0.7907493114471436, 'reward_meter_std': 0.28211304545402527, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7878605723381042, 'reward_repeat_penalty_std': 0.3069226145744324, 'reward_near_duplicate_penalty_mean': 0.9179651141166687, 'reward_near_duplicate_penalty_std': 0.13240814208984375, 'reward_opening_diversity_mean': 0.8999999761581421, 'reward_opening_diversity_std': 0.24494896829128265, 'reward_distinct_2_mean': 0.8616527318954468, 'reward_distinct_2_std': 0.2173728346824646, 'reward_judge_quality_mean': 0.35874998569488525, 'reward_judge_quality_std': 0.2547512948513031, 'reward_total_composite_mean': 0.29524925351142883, 'reward_total_composite_std': 0.27520182728767395, 'epoch': 0.01}

 11%|█▏        | 372/3300 [1:24:00<9:54:48, 12.19s/it]INFO 04-12 12:10:13 [block_pool.py:378] Successfully reset prefix cache

 11%|█▏        | 373/3300 [1:24:15<10:33:55, 12.99s/it]2026-04-12 12:10:27,635 | INFO | train_grpo_train | metrics_logged mode=train step=373

                                                       
{'loss': -0.1954, 'grad_norm': 4.453275680541992, 'learning_rate': 8.872727272727275e-06, 'num_tokens': 745846.0, 'completions/mean_length': 190.25, 'completions/min_length': 121.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 144.2857208251953, 'completions/min_terminated_length': 121.0, 'completions/max_terminated_length': 175.0, 'rewards/meter/mean': 0.7178395986557007, 'rewards/meter/std': 0.31169822812080383, 'rewards/count_adherence/mean': 0.8214285373687744, 'rewards/count_adherence/std': 0.10101524740457535, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9609147310256958, 'rewards/lexical_plausibility/std': 0.11054975539445877, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.053452249616384506, 'rewards/repeat_penalty/mean': 0.2428719401359558, 'rewards/repeat_penalty/std': 0.325005441904068, 'rewards/near_duplicate_penalty/mean': 0.7483271360397339, 'rewards/near_duplicate_penalty/std': 0.13887473940849304, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.0708683431148529, 'rewards/distinct_2/mean': 0.3178720474243164, 'rewards/distinct_2/std': 0.32134172320365906, 'rewards/total_composite/mean': 0.0996420755982399, 'rewards/total_composite/std': 0.06269153207540512, 'reward': 0.0996420755982399, 'reward_std': 0.06269153207540512, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10872436314821243, 'sampling/sampling_logp_difference/max': 2.3243393898010254, 'sampling/importance_sampling_ratio/min': 0.09784805774688721, 'sampling/importance_sampling_ratio/mean': 1.013803243637085, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7211753651499748, 'clip_ratio/low_mean': 0.03993580583482981, 'clip_ratio/low_min': 0.03993580583482981, 'clip_ratio/high_mean': 0.04232870461419225, 'clip_ratio/high_max': 0.04232870461419225, 'clip_ratio/region_mean': 0.08226451044902205, 'reward_total_mean': 0.0996420755982399, 'reward_meter_mean': 0.7178395986557007, 'reward_meter_std': 0.31169822812080383, 'reward_count_adherence_mean': 0.8214285373687744, 'reward_count_adherence_std': 0.10101524740457535, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9609147310256958, 'reward_lexical_plausibility_std': 0.11054975539445877, 'reward_repeat_penalty_mean': 0.2428719401359558, 'reward_repeat_penalty_std': 0.325005441904068, 'reward_near_duplicate_penalty_mean': 0.7483271360397339, 'reward_near_duplicate_penalty_std': 0.13887473940849304, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.0708683431148529, 'reward_distinct_2_mean': 0.3178720474243164, 'reward_distinct_2_std': 0.32134172320365906, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.053452249616384506, 'reward_total_composite_mean': 0.0996420755982399, 'reward_total_composite_std': 0.06269153207540512, 'epoch': 0.01}

 11%|█▏        | 373/3300 [1:24:15<10:33:55, 12.99s/it]INFO 04-12 12:10:27 [block_pool.py:378] Successfully reset prefix cache

 11%|█▏        | 374/3300 [1:24:24<9:35:24, 11.80s/it] 2026-04-12 12:10:36,639 | INFO | train_grpo_train | metrics_logged mode=train step=374

                                                      
{'loss': 0.1091, 'grad_norm': 41.12314224243164, 'learning_rate': 8.86969696969697e-06, 'num_tokens': 747377.0, 'completions/mean_length': 36.375, 'completions/min_length': 31.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.375, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.2555834650993347, 'rewards/meter/std': 0.29624059796333313, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9736520051956177, 'rewards/lexical_plausibility/std': 0.06404155492782593, 'rewards/judge_quality/mean': 0.6387499570846558, 'rewards/judge_quality/std': 0.24718055129051208, 'rewards/repeat_penalty/mean': 0.9591344594955444, 'rewards/repeat_penalty/std': 0.09540197253227234, 'rewards/near_duplicate_penalty/mean': 0.9807000756263733, 'rewards/near_duplicate_penalty/std': 0.035517483949661255, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9759615659713745, 'rewards/distinct_2/std': 0.06799104809761047, 'rewards/total_composite/mean': 0.1528775990009308, 'rewards/total_composite/std': 0.2018582969903946, 'reward': 0.1528775990009308, 'reward_std': 0.2018582969903946, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21097850799560547, 'sampling/sampling_logp_difference/max': 3.235154628753662, 'sampling/importance_sampling_ratio/min': 0.0393541194498539, 'sampling/importance_sampling_ratio/mean': 0.9728639125823975, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.014286793768406, 'clip_ratio/low_mean': 0.12410887982696295, 'clip_ratio/low_min': 0.12410887982696295, 'clip_ratio/high_mean': 0.04423564113676548, 'clip_ratio/high_max': 0.04423564113676548, 'clip_ratio/region_mean': 0.16834452096372843, 'reward_total_mean': 0.1528775990009308, 'reward_meter_mean': 0.2555834650993347, 'reward_meter_std': 0.29624059796333313, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9736520051956177, 'reward_lexical_plausibility_std': 0.06404155492782593, 'reward_repeat_penalty_mean': 0.9591344594955444, 'reward_repeat_penalty_std': 0.09540197253227234, 'reward_near_duplicate_penalty_mean': 0.9807000756263733, 'reward_near_duplicate_penalty_std': 0.035517483949661255, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9759615659713745, 'reward_distinct_2_std': 0.06799104809761047, 'reward_judge_quality_mean': 0.6387499570846558, 'reward_judge_quality_std': 0.24718055129051208, 'reward_total_composite_mean': 0.1528775990009308, 'reward_total_composite_std': 0.2018582969903946, 'epoch': 0.02}

 11%|█▏        | 374/3300 [1:24:24<9:35:24, 11.80s/it]INFO 04-12 12:10:36 [block_pool.py:378] Successfully reset prefix cache

 11%|█▏        | 375/3300 [1:24:39<10:23:03, 12.78s/it]2026-04-12 12:10:51,711 | INFO | train_grpo_train | metrics_logged mode=train step=375

                                                       
{'loss': -0.0953, 'grad_norm': 5.994908332824707, 'learning_rate': 8.866666666666668e-06, 'num_tokens': 749572.0, 'completions/mean_length': 149.375, 'completions/min_length': 81.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 97.5714340209961, 'completions/min_terminated_length': 81.0, 'completions/max_terminated_length': 126.0, 'rewards/meter/mean': 0.5356355905532837, 'rewards/meter/std': 0.4026618003845215, '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/lexical_plausibility/mean': 0.9289287328720093, 'rewards/lexical_plausibility/std': 0.16733857989311218, 'rewards/judge_quality/mean': 0.3375000059604645, 'rewards/judge_quality/std': 0.18077215552330017, 'rewards/repeat_penalty/mean': 0.9829912781715393, 'rewards/repeat_penalty/std': 0.025840558111667633, 'rewards/near_duplicate_penalty/mean': 0.9910508990287781, 'rewards/near_duplicate_penalty/std': 0.005874638445675373, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9918166995048523, 'rewards/distinct_2/std': 0.02314588986337185, 'rewards/total_composite/mean': 0.14762334525585175, 'rewards/total_composite/std': 0.16437840461730957, 'reward': 0.14762334525585175, 'reward_std': 0.16437838971614838, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19908767938613892, 'sampling/sampling_logp_difference/max': 1.4282140731811523, 'sampling/importance_sampling_ratio/min': 0.23973669111728668, 'sampling/importance_sampling_ratio/mean': 1.054633378982544, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2022953629493713, 'clip_ratio/low_mean': 0.09793355409055948, 'clip_ratio/low_min': 0.09793355409055948, 'clip_ratio/high_mean': 0.057337455451488495, 'clip_ratio/high_max': 0.057337455451488495, 'clip_ratio/region_mean': 0.15527100954204798, 'reward_total_mean': 0.14762334525585175, 'reward_meter_mean': 0.5356355905532837, 'reward_meter_std': 0.4026618003845215, '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_lexical_plausibility_mean': 0.9289287328720093, 'reward_lexical_plausibility_std': 0.16733857989311218, 'reward_repeat_penalty_mean': 0.9829912781715393, 'reward_repeat_penalty_std': 0.025840558111667633, 'reward_near_duplicate_penalty_mean': 0.9910508990287781, 'reward_near_duplicate_penalty_std': 0.005874638445675373, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9918166995048523, 'reward_distinct_2_std': 0.02314588986337185, 'reward_judge_quality_mean': 0.3375000059604645, 'reward_judge_quality_std': 0.18077215552330017, 'reward_total_composite_mean': 0.14762334525585175, 'reward_total_composite_std': 0.16437840461730957, 'epoch': 0.02}

 11%|█▏        | 375/3300 [1:24:39<10:23:03, 12.78s/it]INFO 04-12 12:10:51 [block_pool.py:378] Successfully reset prefix cache

 11%|█▏        | 376/3300 [1:24:48<9:26:21, 11.62s/it] 2026-04-12 12:11:00,627 | INFO | train_grpo_train | metrics_logged mode=train step=376

                                                      
{'loss': 0.0492, 'grad_norm': 14.439682960510254, 'learning_rate': 8.863636363636365e-06, 'num_tokens': 751219.0, 'completions/mean_length': 42.875, 'completions/min_length': 35.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.875, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.8486638069152832, 'rewards/meter/std': 0.3000448942184448, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6050000190734863, 'rewards/judge_quality/std': 0.2562365233898163, 'rewards/repeat_penalty/mean': 0.9722264409065247, 'rewards/repeat_penalty/std': 0.032479945570230484, 'rewards/near_duplicate_penalty/mean': 0.9722264409065247, 'rewards/near_duplicate_penalty/std': 0.032479945570230484, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5369139909744263, 'rewards/total_composite/std': 0.3224915564060211, 'reward': 0.5369139909744263, 'reward_std': 0.3224915564060211, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19373628497123718, 'sampling/sampling_logp_difference/max': 1.5680522918701172, 'sampling/importance_sampling_ratio/min': 0.2084507793188095, 'sampling/importance_sampling_ratio/mean': 1.0286318063735962, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.440964788198471, 'clip_ratio/low_mean': 0.07526303362101316, 'clip_ratio/low_min': 0.07526303362101316, 'clip_ratio/high_mean': 0.09555895812809467, 'clip_ratio/high_max': 0.09555895812809467, 'clip_ratio/region_mean': 0.17082199174910784, 'reward_total_mean': 0.5369139909744263, 'reward_meter_mean': 0.8486638069152832, 'reward_meter_std': 0.3000448942184448, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9722264409065247, 'reward_repeat_penalty_std': 0.032479945570230484, 'reward_near_duplicate_penalty_mean': 0.9722264409065247, 'reward_near_duplicate_penalty_std': 0.032479945570230484, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6050000190734863, 'reward_judge_quality_std': 0.2562365233898163, 'reward_total_composite_mean': 0.5369139909744263, 'reward_total_composite_std': 0.3224915564060211, 'epoch': 0.02}

 11%|█▏        | 376/3300 [1:24:48<9:26:21, 11.62s/it]INFO 04-12 12:11:00 [block_pool.py:378] Successfully reset prefix cache

 11%|█▏        | 377/3300 [1:24:57<8:49:18, 10.87s/it]2026-04-12 12:11:09,726 | INFO | train_grpo_train | metrics_logged mode=train step=377

                                                      
{'loss': 0.0307, 'grad_norm': 14.737468719482422, 'learning_rate': 8.860606060606062e-06, 'num_tokens': 752985.0, 'completions/mean_length': 55.75, 'completions/min_length': 45.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.75, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.6870785355567932, 'rewards/meter/std': 0.3609475791454315, '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/lexical_plausibility/mean': 0.9819004535675049, 'rewards/lexical_plausibility/std': 0.051193248480558395, 'rewards/judge_quality/mean': 0.42499998211860657, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.9524993896484375, 'rewards/repeat_penalty/std': 0.03423053398728371, 'rewards/near_duplicate_penalty/mean': 0.9809752702713013, 'rewards/near_duplicate_penalty/std': 0.01396517176181078, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.970942497253418, 'rewards/distinct_2/std': 0.03109455294907093, 'rewards/total_composite/mean': 0.27244529128074646, 'rewards/total_composite/std': 0.198111429810524, 'reward': 0.27244529128074646, 'reward_std': 0.1981114149093628, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18403708934783936, 'sampling/sampling_logp_difference/max': 1.1885719299316406, 'sampling/importance_sampling_ratio/min': 0.3046560287475586, 'sampling/importance_sampling_ratio/mean': 1.0245411396026611, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4230965971946716, 'clip_ratio/low_mean': 0.06366641912609339, 'clip_ratio/low_min': 0.06366641912609339, 'clip_ratio/high_mean': 0.11802011355757713, 'clip_ratio/high_max': 0.11802011355757713, 'clip_ratio/region_mean': 0.18168653268367052, 'reward_total_mean': 0.27244529128074646, 'reward_meter_mean': 0.6870785355567932, 'reward_meter_std': 0.3609475791454315, '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_lexical_plausibility_mean': 0.9819004535675049, 'reward_lexical_plausibility_std': 0.051193248480558395, 'reward_repeat_penalty_mean': 0.9524993896484375, 'reward_repeat_penalty_std': 0.03423053398728371, 'reward_near_duplicate_penalty_mean': 0.9809752702713013, 'reward_near_duplicate_penalty_std': 0.01396517176181078, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.970942497253418, 'reward_distinct_2_std': 0.03109455294907093, 'reward_judge_quality_mean': 0.42499998211860657, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.27244529128074646, 'reward_total_composite_std': 0.198111429810524, 'epoch': 0.02}

 11%|█▏        | 377/3300 [1:24:57<8:49:18, 10.87s/it]INFO 04-12 12:11:09 [block_pool.py:378] Successfully reset prefix cache

 11%|█▏        | 378/3300 [1:25:05<8:10:10, 10.07s/it]2026-04-12 12:11:17,926 | INFO | train_grpo_train | metrics_logged mode=train step=378

                                                      
{'loss': 0.0757, 'grad_norm': 10.974468231201172, 'learning_rate': 8.857575757575758e-06, 'num_tokens': 754475.0, 'completions/mean_length': 38.25, 'completions/min_length': 28.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.25, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.8403074145317078, 'rewards/meter/std': 0.29901474714279175, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.9424999952316284, 'rewards/judge_quality/std': 0.013887288980185986, 'rewards/repeat_penalty/mean': 0.9988548159599304, 'rewards/repeat_penalty/std': 0.0021204412914812565, 'rewards/near_duplicate_penalty/mean': 0.9988548159599304, 'rewards/near_duplicate_penalty/std': 0.0021204412914812565, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.7942900657653809, 'rewards/total_composite/std': 0.28384438157081604, 'reward': 0.7942900657653809, 'reward_std': 0.28384435176849365, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09076273441314697, 'sampling/sampling_logp_difference/max': 1.290827751159668, 'sampling/importance_sampling_ratio/min': 0.2750430107116699, 'sampling/importance_sampling_ratio/mean': 1.006481647491455, 'sampling/importance_sampling_ratio/max': 1.91322660446167, 'entropy': 0.5385432429611683, 'clip_ratio/low_mean': 0.023403679952025414, 'clip_ratio/low_min': 0.023403679952025414, 'clip_ratio/high_mean': 0.07889919588342309, 'clip_ratio/high_max': 0.07889919588342309, 'clip_ratio/region_mean': 0.1023028758354485, 'reward_total_mean': 0.7942900657653809, 'reward_meter_mean': 0.8403074145317078, 'reward_meter_std': 0.29901474714279175, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9988548159599304, 'reward_repeat_penalty_std': 0.0021204412914812565, 'reward_near_duplicate_penalty_mean': 0.9988548159599304, 'reward_near_duplicate_penalty_std': 0.0021204412914812565, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.9424999952316284, 'reward_judge_quality_std': 0.013887288980185986, 'reward_total_composite_mean': 0.7942900657653809, 'reward_total_composite_std': 0.28384438157081604, 'epoch': 0.02}

 11%|█▏        | 378/3300 [1:25:05<8:10:10, 10.07s/it]INFO 04-12 12:11:18 [block_pool.py:378] Successfully reset prefix cache

 11%|█▏        | 379/3300 [1:25:14<7:47:10,  9.60s/it]2026-04-12 12:11:26,428 | INFO | train_grpo_train | metrics_logged mode=train step=379

                                                      
{'loss': 0.0397, 'grad_norm': 18.5172119140625, 'learning_rate': 8.854545454545455e-06, 'num_tokens': 755762.0, 'completions/mean_length': 21.875, 'completions/min_length': 17.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.875, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.6731079816818237, 'rewards/meter/std': 0.4193631112575531, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9007352590560913, 'rewards/lexical_plausibility/std': 0.13948148488998413, 'rewards/judge_quality/mean': 0.7549999952316284, 'rewards/judge_quality/std': 0.25286927819252014, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5350714325904846, 'rewards/total_composite/std': 0.4054333567619324, 'reward': 0.5350714325904846, 'reward_std': 0.40543332695961, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1621473729610443, 'sampling/sampling_logp_difference/max': 1.5050764083862305, 'sampling/importance_sampling_ratio/min': 0.2220003455877304, 'sampling/importance_sampling_ratio/mean': 1.0216091871261597, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3622380644083023, 'clip_ratio/low_mean': 0.1026265425607562, 'clip_ratio/low_min': 0.1026265425607562, 'clip_ratio/high_mean': 0.05757147399708629, 'clip_ratio/high_max': 0.05757147399708629, 'clip_ratio/region_mean': 0.1601980165578425, 'reward_total_mean': 0.5350714325904846, 'reward_meter_mean': 0.6731079816818237, 'reward_meter_std': 0.4193631112575531, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9007352590560913, 'reward_lexical_plausibility_std': 0.13948148488998413, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7549999952316284, 'reward_judge_quality_std': 0.25286927819252014, 'reward_total_composite_mean': 0.5350714325904846, 'reward_total_composite_std': 0.4054333567619324, 'epoch': 0.02}

 11%|█▏        | 379/3300 [1:25:14<7:47:10,  9.60s/it]INFO 04-12 12:11:26 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 380/3300 [1:25:25<8:14:17, 10.16s/it]2026-04-12 12:11:37,893 | INFO | train_grpo_train | metrics_logged mode=train step=380

                                                      
{'loss': 0.0575, 'grad_norm': 11.882780075073242, 'learning_rate': 8.851515151515152e-06, 'num_tokens': 757708.0, 'completions/mean_length': 79.25, 'completions/min_length': 64.0, 'completions/max_length': 100.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.25, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 100.0, 'rewards/meter/mean': 0.6264046430587769, 'rewards/meter/std': 0.3158259689807892, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.1060660183429718, 'rewards/repeat_penalty/mean': 0.8990840315818787, 'rewards/repeat_penalty/std': 0.11603206396102905, 'rewards/near_duplicate_penalty/mean': 0.9785974025726318, 'rewards/near_duplicate_penalty/std': 0.00721503421664238, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9187901020050049, 'rewards/distinct_2/std': 0.11846604943275452, 'rewards/total_composite/mean': 0.16143256425857544, 'rewards/total_composite/std': 0.1607261598110199, 'reward': 0.16143256425857544, 'reward_std': 0.1607261449098587, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18200050294399261, 'sampling/sampling_logp_difference/max': 1.4343328475952148, 'sampling/importance_sampling_ratio/min': 0.23827427625656128, 'sampling/importance_sampling_ratio/mean': 1.0386003255844116, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3949623852968216, 'clip_ratio/low_mean': 0.11346772685647011, 'clip_ratio/low_min': 0.11346772685647011, 'clip_ratio/high_mean': 0.062230393290519714, 'clip_ratio/high_max': 0.062230393290519714, 'clip_ratio/region_mean': 0.17569812014698982, 'reward_total_mean': 0.16143256425857544, 'reward_meter_mean': 0.6264046430587769, 'reward_meter_std': 0.3158259689807892, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8990840315818787, 'reward_repeat_penalty_std': 0.11603206396102905, 'reward_near_duplicate_penalty_mean': 0.9785974025726318, 'reward_near_duplicate_penalty_std': 0.00721503421664238, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9187901020050049, 'reward_distinct_2_std': 0.11846604943275452, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.1060660183429718, 'reward_total_composite_mean': 0.16143256425857544, 'reward_total_composite_std': 0.1607261598110199, 'epoch': 0.02}

 12%|█▏        | 380/3300 [1:25:25<8:14:17, 10.16s/it]INFO 04-12 12:11:38 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 381/3300 [1:25:40<9:17:43, 11.46s/it]2026-04-12 12:11:52,408 | INFO | train_grpo_train | metrics_logged mode=train step=381

                                                      
{'loss': -0.0713, 'grad_norm': 5.488112926483154, 'learning_rate': 8.84848484848485e-06, 'num_tokens': 759306.0, 'completions/mean_length': 101.75, 'completions/min_length': 24.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 43.142860412597656, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.5872917175292969, 'rewards/meter/std': 0.45493966341018677, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9488881230354309, 'rewards/lexical_plausibility/std': 0.14456620812416077, 'rewards/judge_quality/mean': 0.5249999761581421, 'rewards/judge_quality/std': 0.30019041895866394, 'rewards/repeat_penalty/mean': 0.929169774055481, 'rewards/repeat_penalty/std': 0.1116023063659668, 'rewards/near_duplicate_penalty/mean': 0.991669774055481, 'rewards/near_duplicate_penalty/std': 0.015882477164268494, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3641815185546875, 'rewards/total_composite/std': 0.36358609795570374, 'reward': 0.3641815185546875, 'reward_std': 0.36358609795570374, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17417752742767334, 'sampling/sampling_logp_difference/max': 1.2228634357452393, 'sampling/importance_sampling_ratio/min': 0.29438599944114685, 'sampling/importance_sampling_ratio/mean': 1.01909601688385, 'sampling/importance_sampling_ratio/max': 1.985019326210022, 'entropy': 1.4191492050886154, 'clip_ratio/low_mean': 0.07792874425649643, 'clip_ratio/low_min': 0.07792874425649643, 'clip_ratio/high_mean': 0.09363781986758113, 'clip_ratio/high_max': 0.09363781986758113, 'clip_ratio/region_mean': 0.17156656412407756, 'reward_total_mean': 0.3641815185546875, 'reward_meter_mean': 0.5872917175292969, 'reward_meter_std': 0.45493966341018677, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9488881230354309, 'reward_lexical_plausibility_std': 0.14456620812416077, 'reward_repeat_penalty_mean': 0.929169774055481, 'reward_repeat_penalty_std': 0.1116023063659668, 'reward_near_duplicate_penalty_mean': 0.991669774055481, 'reward_near_duplicate_penalty_std': 0.015882477164268494, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5249999761581421, 'reward_judge_quality_std': 0.30019041895866394, 'reward_total_composite_mean': 0.3641815185546875, 'reward_total_composite_std': 0.36358609795570374, 'epoch': 0.02}

 12%|█▏        | 381/3300 [1:25:40<9:17:43, 11.46s/it]INFO 04-12 12:11:52 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 382/3300 [1:25:54<9:49:39, 12.12s/it]2026-04-12 12:12:06,072 | INFO | train_grpo_train | metrics_logged mode=train step=382

                                                      
{'loss': -0.0402, 'grad_norm': 2.7200467586517334, 'learning_rate': 8.845454545454547e-06, 'num_tokens': 760702.0, 'completions/mean_length': 155.5, 'completions/min_length': 31.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 36.66666793823242, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.4661267101764679, 'rewards/meter/std': 0.4470050632953644, '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/lexical_plausibility/mean': 0.8494057655334473, 'rewards/lexical_plausibility/std': 0.2792769968509674, 'rewards/judge_quality/mean': 0.6312500238418579, 'rewards/judge_quality/std': 0.40912583470344543, 'rewards/repeat_penalty/mean': 0.9924811124801636, 'rewards/repeat_penalty/std': 0.010740749537944794, 'rewards/near_duplicate_penalty/mean': 0.9924811124801636, 'rewards/near_duplicate_penalty/std': 0.010740749537944794, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.35822466015815735, 'rewards/total_composite/std': 0.37213534116744995, 'reward': 0.35822466015815735, 'reward_std': 0.37213531136512756, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1962714046239853, 'sampling/sampling_logp_difference/max': 1.5085039138793945, 'sampling/importance_sampling_ratio/min': 0.22124072909355164, 'sampling/importance_sampling_ratio/mean': 1.04746675491333, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4118845611810684, 'clip_ratio/low_mean': 0.062454101629555225, 'clip_ratio/low_min': 0.062454101629555225, 'clip_ratio/high_mean': 0.08005218207836151, 'clip_ratio/high_max': 0.08005218207836151, 'clip_ratio/region_mean': 0.14250628370791674, 'reward_total_mean': 0.35822466015815735, 'reward_meter_mean': 0.4661267101764679, 'reward_meter_std': 0.4470050632953644, '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_lexical_plausibility_mean': 0.8494057655334473, 'reward_lexical_plausibility_std': 0.2792769968509674, 'reward_repeat_penalty_mean': 0.9924811124801636, 'reward_repeat_penalty_std': 0.010740749537944794, 'reward_near_duplicate_penalty_mean': 0.9924811124801636, 'reward_near_duplicate_penalty_std': 0.010740749537944794, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6312500238418579, 'reward_judge_quality_std': 0.40912583470344543, 'reward_total_composite_mean': 0.35822466015815735, 'reward_total_composite_std': 0.37213534116744995, 'epoch': 0.02}

 12%|█▏        | 382/3300 [1:25:54<9:49:39, 12.12s/it]INFO 04-12 12:12:06 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 383/3300 [1:26:08<10:26:17, 12.88s/it]2026-04-12 12:12:20,723 | INFO | train_grpo_train | metrics_logged mode=train step=383

                                                       
{'loss': -0.1226, 'grad_norm': 3.965210199356079, 'learning_rate': 8.842424242424244e-06, 'num_tokens': 763085.0, 'completions/mean_length': 216.875, 'completions/min_length': 114.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 118.5, 'completions/min_terminated_length': 114.0, 'completions/max_terminated_length': 124.0, 'rewards/meter/mean': 0.8149964809417725, 'rewards/meter/std': 0.2945411205291748, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9053611159324646, 'rewards/lexical_plausibility/std': 0.186506986618042, 'rewards/judge_quality/mean': 0.35374999046325684, 'rewards/judge_quality/std': 0.25933638215065, 'rewards/repeat_penalty/mean': 0.9663739800453186, 'rewards/repeat_penalty/std': 0.0157057773321867, 'rewards/near_duplicate_penalty/mean': 0.9774044752120972, 'rewards/near_duplicate_penalty/std': 0.014795432798564434, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9887775182723999, 'rewards/distinct_2/std': 0.013228368945419788, 'rewards/total_composite/mean': 0.30963075160980225, 'rewards/total_composite/std': 0.29092901945114136, 'reward': 0.30963075160980225, 'reward_std': 0.29092898964881897, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17502562701702118, 'sampling/sampling_logp_difference/max': 1.551018238067627, 'sampling/importance_sampling_ratio/min': 0.21203197538852692, 'sampling/importance_sampling_ratio/mean': 1.0135414600372314, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0135677829384804, 'clip_ratio/low_mean': 0.013888888992369175, 'clip_ratio/low_min': 0.013888888992369175, 'clip_ratio/high_mean': 0.10714712552726269, 'clip_ratio/high_max': 0.10714712552726269, 'clip_ratio/region_mean': 0.12103601451963186, 'reward_total_mean': 0.30963075160980225, 'reward_meter_mean': 0.8149964809417725, 'reward_meter_std': 0.2945411205291748, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9053611159324646, 'reward_lexical_plausibility_std': 0.186506986618042, 'reward_repeat_penalty_mean': 0.9663739800453186, 'reward_repeat_penalty_std': 0.0157057773321867, 'reward_near_duplicate_penalty_mean': 0.9774044752120972, 'reward_near_duplicate_penalty_std': 0.014795432798564434, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9887775182723999, 'reward_distinct_2_std': 0.013228368945419788, 'reward_judge_quality_mean': 0.35374999046325684, 'reward_judge_quality_std': 0.25933638215065, 'reward_total_composite_mean': 0.30963075160980225, 'reward_total_composite_std': 0.29092901945114136, 'epoch': 0.02}

 12%|█▏        | 383/3300 [1:26:08<10:26:17, 12.88s/it]INFO 04-12 12:12:20 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 384/3300 [1:26:17<9:26:19, 11.65s/it] 2026-04-12 12:12:29,508 | INFO | train_grpo_train | metrics_logged mode=train step=384

                                                      
{'loss': 0.0435, 'grad_norm': 12.701199531555176, 'learning_rate': 8.83939393939394e-06, 'num_tokens': 764932.0, 'completions/mean_length': 49.875, 'completions/min_length': 44.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.875, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.6360716819763184, 'rewards/meter/std': 0.4111797511577606, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9864082336425781, 'rewards/lexical_plausibility/std': 0.0384434312582016, 'rewards/judge_quality/mean': 0.7850000262260437, 'rewards/judge_quality/std': 0.26880690455436707, 'rewards/repeat_penalty/mean': 0.9830220937728882, 'rewards/repeat_penalty/std': 0.022771036252379417, 'rewards/near_duplicate_penalty/mean': 0.9913833141326904, 'rewards/near_duplicate_penalty/std': 0.011131873354315758, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9916387796401978, 'rewards/distinct_2/std': 0.0236490610986948, 'rewards/total_composite/mean': 0.5121535062789917, 'rewards/total_composite/std': 0.3908560574054718, 'reward': 0.5121535062789917, 'reward_std': 0.3908560574054718, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13641001284122467, 'sampling/sampling_logp_difference/max': 1.4167276620864868, 'sampling/importance_sampling_ratio/min': 0.24250628054141998, 'sampling/importance_sampling_ratio/mean': 1.0251556634902954, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0516512989997864, 'clip_ratio/low_mean': 0.054174209013581276, 'clip_ratio/low_min': 0.054174209013581276, 'clip_ratio/high_mean': 0.057848403230309486, 'clip_ratio/high_max': 0.057848403230309486, 'clip_ratio/region_mean': 0.11202261224389076, 'reward_total_mean': 0.5121535062789917, 'reward_meter_mean': 0.6360716819763184, 'reward_meter_std': 0.4111797511577606, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9864082336425781, 'reward_lexical_plausibility_std': 0.0384434312582016, 'reward_repeat_penalty_mean': 0.9830220937728882, 'reward_repeat_penalty_std': 0.022771036252379417, 'reward_near_duplicate_penalty_mean': 0.9913833141326904, 'reward_near_duplicate_penalty_std': 0.011131873354315758, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9916387796401978, 'reward_distinct_2_std': 0.0236490610986948, 'reward_judge_quality_mean': 0.7850000262260437, 'reward_judge_quality_std': 0.26880690455436707, 'reward_total_composite_mean': 0.5121535062789917, 'reward_total_composite_std': 0.3908560574054718, 'epoch': 0.02}

 12%|█▏        | 384/3300 [1:26:17<9:26:19, 11.65s/it]INFO 04-12 12:12:29 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 385/3300 [1:26:32<10:11:29, 12.59s/it]2026-04-12 12:12:44,274 | INFO | train_grpo_train | metrics_logged mode=train step=385

                                                       
{'loss': -0.0398, 'grad_norm': 4.453365325927734, 'learning_rate': 8.836363636363637e-06, 'num_tokens': 766536.0, 'completions/mean_length': 87.5, 'completions/min_length': 24.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 26.85714340209961, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.5383661985397339, 'rewards/meter/std': 0.4269161522388458, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.7755584716796875, 'rewards/lexical_plausibility/std': 0.22294670343399048, 'rewards/judge_quality/mean': 0.7437500357627869, 'rewards/judge_quality/std': 0.3452095687389374, 'rewards/repeat_penalty/mean': 0.78125, 'rewards/repeat_penalty/std': 0.0883883461356163, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.42758840322494507, 'rewards/total_composite/std': 0.3589273989200592, 'reward': 0.42758840322494507, 'reward_std': 0.3589273989200592, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16211320459842682, 'sampling/sampling_logp_difference/max': 1.229848861694336, 'sampling/importance_sampling_ratio/min': 0.29233676195144653, 'sampling/importance_sampling_ratio/mean': 1.0205551385879517, 'sampling/importance_sampling_ratio/max': 1.9664716720581055, 'entropy': 1.1866639107465744, 'clip_ratio/low_mean': 0.03215939225628972, 'clip_ratio/low_min': 0.03215939225628972, 'clip_ratio/high_mean': 0.08019467163830996, 'clip_ratio/high_max': 0.08019467163830996, 'clip_ratio/region_mean': 0.11235406389459968, 'reward_total_mean': 0.42758840322494507, 'reward_meter_mean': 0.5383661985397339, 'reward_meter_std': 0.4269161522388458, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.7755584716796875, 'reward_lexical_plausibility_std': 0.22294670343399048, 'reward_repeat_penalty_mean': 0.78125, 'reward_repeat_penalty_std': 0.0883883461356163, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7437500357627869, 'reward_judge_quality_std': 0.3452095687389374, 'reward_total_composite_mean': 0.42758840322494507, 'reward_total_composite_std': 0.3589273989200592, 'epoch': 0.02}

 12%|█▏        | 385/3300 [1:26:32<10:11:29, 12.59s/it]INFO 04-12 12:12:44 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 386/3300 [1:26:42<9:37:10, 11.88s/it] 2026-04-12 12:12:54,519 | INFO | train_grpo_train | metrics_logged mode=train step=386

                                                      
{'loss': -0.0063, 'grad_norm': 13.448062896728516, 'learning_rate': 8.833333333333334e-06, 'num_tokens': 768075.0, 'completions/mean_length': 39.375, 'completions/min_length': 36.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.375, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.8317387700080872, 'rewards/meter/std': 0.20346671342849731, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9910990595817566, 'rewards/lexical_plausibility/std': 0.02517561987042427, 'rewards/judge_quality/mean': 0.6637499928474426, 'rewards/judge_quality/std': 0.28465205430984497, 'rewards/repeat_penalty/mean': 0.9798574447631836, 'rewards/repeat_penalty/std': 0.041568104177713394, 'rewards/near_duplicate_penalty/mean': 0.9901480674743652, 'rewards/near_duplicate_penalty/std': 0.015714745968580246, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9893162250518799, 'rewards/distinct_2/std': 0.03021823801100254, 'rewards/total_composite/mean': 0.5324898958206177, 'rewards/total_composite/std': 0.2541130781173706, 'reward': 0.5324898958206177, 'reward_std': 0.2541130781173706, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1548158973455429, 'sampling/sampling_logp_difference/max': 1.1048154830932617, 'sampling/importance_sampling_ratio/min': 0.38023486733436584, 'sampling/importance_sampling_ratio/mean': 1.0520660877227783, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3426322862505913, 'clip_ratio/low_mean': 0.0852726970333606, 'clip_ratio/low_min': 0.0852726970333606, 'clip_ratio/high_mean': 0.04587095510214567, 'clip_ratio/high_max': 0.04587095510214567, 'clip_ratio/region_mean': 0.13114365213550627, 'reward_total_mean': 0.5324898958206177, 'reward_meter_mean': 0.8317387700080872, 'reward_meter_std': 0.20346671342849731, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9910990595817566, 'reward_lexical_plausibility_std': 0.02517561987042427, 'reward_repeat_penalty_mean': 0.9798574447631836, 'reward_repeat_penalty_std': 0.041568104177713394, 'reward_near_duplicate_penalty_mean': 0.9901480674743652, 'reward_near_duplicate_penalty_std': 0.015714745968580246, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9893162250518799, 'reward_distinct_2_std': 0.03021823801100254, 'reward_judge_quality_mean': 0.6637499928474426, 'reward_judge_quality_std': 0.28465205430984497, 'reward_total_composite_mean': 0.5324898958206177, 'reward_total_composite_std': 0.2541130781173706, 'epoch': 0.02}

 12%|█▏        | 386/3300 [1:26:42<9:37:10, 11.88s/it]INFO 04-12 12:12:54 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 387/3300 [1:26:56<10:05:25, 12.47s/it]2026-04-12 12:13:08,354 | INFO | train_grpo_train | metrics_logged mode=train step=387

                                                       
{'loss': -0.0227, 'grad_norm': 4.439444065093994, 'learning_rate': 8.830303030303031e-06, 'num_tokens': 769497.0, 'completions/mean_length': 88.75, 'completions/min_length': 23.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 28.285715103149414, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.8510254621505737, 'rewards/meter/std': 0.34386804699897766, '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/lexical_plausibility/mean': 0.8292478322982788, 'rewards/lexical_plausibility/std': 0.26898396015167236, 'rewards/judge_quality/mean': 0.39625000953674316, 'rewards/judge_quality/std': 0.34801220893859863, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3792494535446167, 'rewards/total_composite/std': 0.34440329670906067, 'reward': 0.3792494535446167, 'reward_std': 0.3444032669067383, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2064886838197708, 'sampling/sampling_logp_difference/max': 1.001784324645996, 'sampling/importance_sampling_ratio/min': 0.3672236204147339, 'sampling/importance_sampling_ratio/mean': 1.0405688285827637, 'sampling/importance_sampling_ratio/max': 1.8839207887649536, 'entropy': 1.7930859327316284, 'clip_ratio/low_mean': 0.13832499645650387, 'clip_ratio/low_min': 0.13832499645650387, 'clip_ratio/high_mean': 0.04479166679084301, 'clip_ratio/high_max': 0.04479166679084301, 'clip_ratio/region_mean': 0.18311666324734688, 'reward_total_mean': 0.3792494535446167, 'reward_meter_mean': 0.8510254621505737, 'reward_meter_std': 0.34386804699897766, '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_lexical_plausibility_mean': 0.8292478322982788, 'reward_lexical_plausibility_std': 0.26898396015167236, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.39625000953674316, 'reward_judge_quality_std': 0.34801220893859863, 'reward_total_composite_mean': 0.3792494535446167, 'reward_total_composite_std': 0.34440329670906067, 'epoch': 0.02}

 12%|█▏        | 387/3300 [1:26:56<10:05:25, 12.47s/it]INFO 04-12 12:13:08 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 388/3300 [1:27:05<9:17:12, 11.48s/it] 2026-04-12 12:13:17,530 | INFO | train_grpo_train | metrics_logged mode=train step=388

                                                      
{'loss': 0.0682, 'grad_norm': 13.040472984313965, 'learning_rate': 8.827272727272727e-06, 'num_tokens': 771251.0, 'completions/mean_length': 60.25, 'completions/min_length': 53.0, 'completions/max_length': 82.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.25, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.5014908313751221, 'rewards/meter/std': 0.36363595724105835, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.528749942779541, 'rewards/judge_quality/std': 0.15824371576309204, 'rewards/repeat_penalty/mean': 0.9748153686523438, 'rewards/repeat_penalty/std': 0.03493955358862877, 'rewards/near_duplicate_penalty/mean': 0.9871870279312134, 'rewards/near_duplicate_penalty/std': 0.015612704679369926, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9872616529464722, 'rewards/distinct_2/std': 0.0239923894405365, 'rewards/total_composite/mean': 0.24782469868659973, 'rewards/total_composite/std': 0.1842014640569687, 'reward': 0.24782469868659973, 'reward_std': 0.1842014491558075, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1843993216753006, 'sampling/sampling_logp_difference/max': 1.505722999572754, 'sampling/importance_sampling_ratio/min': 0.22185683250427246, 'sampling/importance_sampling_ratio/mean': 1.0240330696105957, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1352952420711517, 'clip_ratio/low_mean': 0.0759481405839324, 'clip_ratio/low_min': 0.0759481405839324, 'clip_ratio/high_mean': 0.06730315182358027, 'clip_ratio/high_max': 0.06730315182358027, 'clip_ratio/region_mean': 0.14325129240751266, 'reward_total_mean': 0.24782469868659973, 'reward_meter_mean': 0.5014908313751221, 'reward_meter_std': 0.36363595724105835, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9748153686523438, 'reward_repeat_penalty_std': 0.03493955358862877, 'reward_near_duplicate_penalty_mean': 0.9871870279312134, 'reward_near_duplicate_penalty_std': 0.015612704679369926, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9872616529464722, 'reward_distinct_2_std': 0.0239923894405365, 'reward_judge_quality_mean': 0.528749942779541, 'reward_judge_quality_std': 0.15824371576309204, 'reward_total_composite_mean': 0.24782469868659973, 'reward_total_composite_std': 0.1842014640569687, 'epoch': 0.02}

 12%|█▏        | 388/3300 [1:27:05<9:17:12, 11.48s/it]INFO 04-12 12:13:17 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 389/3300 [1:27:13<8:24:01, 10.39s/it]2026-04-12 12:13:25,373 | INFO | train_grpo_train | metrics_logged mode=train step=389

                                                      
{'loss': 0.0988, 'grad_norm': 24.313623428344727, 'learning_rate': 8.824242424242426e-06, 'num_tokens': 772631.0, 'completions/mean_length': 19.5, 'completions/min_length': 17.0, 'completions/max_length': 23.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.5, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.804509162902832, 'rewards/meter/std': 0.3564189076423645, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9361213445663452, 'rewards/lexical_plausibility/std': 0.1569535881280899, 'rewards/judge_quality/mean': 0.9200000166893005, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.7480332255363464, 'rewards/repeat_penalty/std': 0.005562899634242058, 'rewards/near_duplicate_penalty/mean': 0.9973776340484619, 'rewards/near_duplicate_penalty/std': 0.007417213171720505, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.7401484847068787, 'rewards/total_composite/std': 0.3279053866863251, 'reward': 0.7401484847068787, 'reward_std': 0.3279053866863251, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1738765388727188, 'sampling/sampling_logp_difference/max': 1.1999459266662598, 'sampling/importance_sampling_ratio/min': 0.30121049284935, 'sampling/importance_sampling_ratio/mean': 1.0502216815948486, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3886692970991135, 'clip_ratio/low_mean': 0.04375000111758709, 'clip_ratio/low_min': 0.04375000111758709, 'clip_ratio/high_mean': 0.10645459638908505, 'clip_ratio/high_max': 0.10645459638908505, 'clip_ratio/region_mean': 0.15020459750667214, 'reward_total_mean': 0.7401484847068787, 'reward_meter_mean': 0.804509162902832, 'reward_meter_std': 0.3564189076423645, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9361213445663452, 'reward_lexical_plausibility_std': 0.1569535881280899, 'reward_repeat_penalty_mean': 0.7480332255363464, 'reward_repeat_penalty_std': 0.005562899634242058, 'reward_near_duplicate_penalty_mean': 0.9973776340484619, 'reward_near_duplicate_penalty_std': 0.007417213171720505, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.9200000166893005, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.7401484847068787, 'reward_total_composite_std': 0.3279053866863251, 'epoch': 0.02}

 12%|█▏        | 389/3300 [1:27:13<8:24:01, 10.39s/it]INFO 04-12 12:13:25 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 390/3300 [1:27:27<9:19:21, 11.53s/it]2026-04-12 12:13:39,571 | INFO | train_grpo_train | metrics_logged mode=train step=390

                                                      
{'loss': -0.0391, 'grad_norm': 2.26735520362854, 'learning_rate': 8.821212121212121e-06, 'num_tokens': 774335.0, 'completions/mean_length': 295.0, 'completions/min_length': 67.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 78.0, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.42222124338150024, 'rewards/meter/std': 0.4066350758075714, 'rewards/count_adherence/mean': 0.7083333730697632, 'rewards/count_adherence/std': 0.37533053755760193, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/lexical_plausibility/mean': 0.601567804813385, 'rewards/lexical_plausibility/std': 0.3263372778892517, 'rewards/judge_quality/mean': 0.29249998927116394, 'rewards/judge_quality/std': 0.38975268602371216, 'rewards/repeat_penalty/mean': 0.9427352547645569, 'rewards/repeat_penalty/std': 0.09331945329904556, 'rewards/near_duplicate_penalty/mean': 0.9914441108703613, 'rewards/near_duplicate_penalty/std': 0.020643942058086395, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9818196296691895, 'rewards/distinct_2/std': 0.03497876599431038, 'rewards/total_composite/mean': 0.14455711841583252, 'rewards/total_composite/std': 0.3151821792125702, 'reward': 0.14455711841583252, 'reward_std': 0.3151821494102478, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21467247605323792, 'sampling/sampling_logp_difference/max': 1.9102625846862793, 'sampling/importance_sampling_ratio/min': 0.1480415165424347, 'sampling/importance_sampling_ratio/mean': 1.051853060722351, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.929746150970459, 'clip_ratio/low_mean': 0.06571061722934246, 'clip_ratio/low_min': 0.06571061722934246, 'clip_ratio/high_mean': 0.020522387698292732, 'clip_ratio/high_max': 0.020522387698292732, 'clip_ratio/region_mean': 0.08623300492763519, 'reward_total_mean': 0.14455711841583252, 'reward_meter_mean': 0.42222124338150024, 'reward_meter_std': 0.4066350758075714, 'reward_count_adherence_mean': 0.7083333730697632, 'reward_count_adherence_std': 0.37533053755760193, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_lexical_plausibility_mean': 0.601567804813385, 'reward_lexical_plausibility_std': 0.3263372778892517, 'reward_repeat_penalty_mean': 0.9427352547645569, 'reward_repeat_penalty_std': 0.09331945329904556, 'reward_near_duplicate_penalty_mean': 0.9914441108703613, 'reward_near_duplicate_penalty_std': 0.020643942058086395, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9818196296691895, 'reward_distinct_2_std': 0.03497876599431038, 'reward_judge_quality_mean': 0.29249998927116394, 'reward_judge_quality_std': 0.38975268602371216, 'reward_total_composite_mean': 0.14455711841583252, 'reward_total_composite_std': 0.3151821792125702, 'epoch': 0.02}

 12%|█▏        | 390/3300 [1:27:27<9:19:21, 11.53s/it]INFO 04-12 12:13:39 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 391/3300 [1:27:41<9:54:39, 12.27s/it]2026-04-12 12:13:53,545 | INFO | train_grpo_train | metrics_logged mode=train step=391

                                                      
{'loss': 0.1176, 'grad_norm': 2.996407985687256, 'learning_rate': 8.818181818181819e-06, 'num_tokens': 775988.0, 'completions/mean_length': 168.625, 'completions/min_length': 19.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 54.16666793823242, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 211.0, 'rewards/meter/mean': 0.7190061807632446, 'rewards/meter/std': 0.44676926732063293, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.6776604056358337, 'rewards/lexical_plausibility/std': 0.39427605271339417, 'rewards/judge_quality/mean': 0.4975000023841858, 'rewards/judge_quality/std': 0.45285600423812866, 'rewards/repeat_penalty/mean': 0.7781673073768616, 'rewards/repeat_penalty/std': 0.0900484099984169, 'rewards/near_duplicate_penalty/mean': 0.9958897829055786, 'rewards/near_duplicate_penalty/std': 0.011625555343925953, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4701194763183594, 'rewards/total_composite/std': 0.4655492305755615, 'reward': 0.4701194763183594, 'reward_std': 0.4655492305755615, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16584241390228271, 'sampling/sampling_logp_difference/max': 1.244649887084961, 'sampling/importance_sampling_ratio/min': 0.28804174065589905, 'sampling/importance_sampling_ratio/mean': 1.035207986831665, 'sampling/importance_sampling_ratio/max': 1.648921012878418, 'entropy': 2.0549472272396088, 'clip_ratio/low_mean': 0.015608902089297771, 'clip_ratio/low_min': 0.015608902089297771, 'clip_ratio/high_mean': 0.0672911899164319, 'clip_ratio/high_max': 0.0672911899164319, 'clip_ratio/region_mean': 0.08290009200572968, 'reward_total_mean': 0.4701194763183594, 'reward_meter_mean': 0.7190061807632446, 'reward_meter_std': 0.44676926732063293, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.6776604056358337, 'reward_lexical_plausibility_std': 0.39427605271339417, 'reward_repeat_penalty_mean': 0.7781673073768616, 'reward_repeat_penalty_std': 0.0900484099984169, 'reward_near_duplicate_penalty_mean': 0.9958897829055786, 'reward_near_duplicate_penalty_std': 0.011625555343925953, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4975000023841858, 'reward_judge_quality_std': 0.45285600423812866, 'reward_total_composite_mean': 0.4701194763183594, 'reward_total_composite_std': 0.4655492305755615, 'epoch': 0.02}

 12%|█▏        | 391/3300 [1:27:41<9:54:39, 12.27s/it]INFO 04-12 12:13:53 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 392/3300 [1:27:51<9:28:04, 11.72s/it]2026-04-12 12:14:03,999 | INFO | train_grpo_train | metrics_logged mode=train step=392

                                                      
{'loss': 0.095, 'grad_norm': 16.90131950378418, 'learning_rate': 8.815151515151516e-06, 'num_tokens': 777546.0, 'completions/mean_length': 42.75, 'completions/min_length': 34.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.75, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.7961781024932861, 'rewards/meter/std': 0.33177298307418823, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5549999475479126, 'rewards/judge_quality/std': 0.22790977358818054, 'rewards/repeat_penalty/mean': 0.9956271648406982, 'rewards/repeat_penalty/std': 0.008242162875831127, 'rewards/near_duplicate_penalty/mean': 0.9956271648406982, 'rewards/near_duplicate_penalty/std': 0.008242162875831127, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4092037081718445, 'rewards/total_composite/std': 0.23701132833957672, 'reward': 0.4092037081718445, 'reward_std': 0.23701131343841553, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20144127309322357, 'sampling/sampling_logp_difference/max': 1.7225580215454102, 'sampling/importance_sampling_ratio/min': 0.17860867083072662, 'sampling/importance_sampling_ratio/mean': 1.0382312536239624, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.946122795343399, 'clip_ratio/low_mean': 0.08899091929197311, 'clip_ratio/low_min': 0.08899091929197311, 'clip_ratio/high_mean': 0.06624577008187771, 'clip_ratio/high_max': 0.06624577008187771, 'clip_ratio/region_mean': 0.15523668937385082, 'reward_total_mean': 0.4092037081718445, 'reward_meter_mean': 0.7961781024932861, 'reward_meter_std': 0.33177298307418823, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9956271648406982, 'reward_repeat_penalty_std': 0.008242162875831127, 'reward_near_duplicate_penalty_mean': 0.9956271648406982, 'reward_near_duplicate_penalty_std': 0.008242162875831127, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5549999475479126, 'reward_judge_quality_std': 0.22790977358818054, 'reward_total_composite_mean': 0.4092037081718445, 'reward_total_composite_std': 0.23701132833957672, 'epoch': 0.02}

 12%|█▏        | 392/3300 [1:27:51<9:28:04, 11.72s/it]INFO 04-12 12:14:04 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 393/3300 [1:28:06<10:11:41, 12.63s/it]2026-04-12 12:14:18,735 | INFO | train_grpo_train | metrics_logged mode=train step=393

                                                       
{'loss': -0.1054, 'grad_norm': 1.9112484455108643, 'learning_rate': 8.812121212121213e-06, 'num_tokens': 779489.0, 'completions/mean_length': 410.875, 'completions/min_length': 106.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.75, 'completions/mean_terminated_length': 107.5, 'completions/min_terminated_length': 106.0, 'completions/max_terminated_length': 109.0, 'rewards/meter/mean': 0.6407234072685242, 'rewards/meter/std': 0.3280641436576843, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.37533053755760193, 'rewards/arabic_clean/mean': 0.25, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.7479768395423889, 'rewards/lexical_plausibility/std': 0.18183663487434387, 'rewards/judge_quality/mean': 0.16250000894069672, 'rewards/judge_quality/std': 0.18077214062213898, 'rewards/repeat_penalty/mean': 0.9357457160949707, 'rewards/repeat_penalty/std': 0.1146545335650444, 'rewards/near_duplicate_penalty/mean': 0.9982457160949707, 'rewards/near_duplicate_penalty/std': 0.0018473644740879536, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.09551124274730682, 'rewards/total_composite/std': 0.17697452008724213, 'reward': 0.09551124274730682, 'reward_std': 0.17697452008724213, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1864493042230606, 'sampling/sampling_logp_difference/max': 1.5367060899734497, 'sampling/importance_sampling_ratio/min': 0.21508842706680298, 'sampling/importance_sampling_ratio/mean': 1.0050638914108276, 'sampling/importance_sampling_ratio/max': 1.7277193069458008, 'entropy': 0.5865364372730255, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.05926519073545933, 'clip_ratio/high_max': 0.05926519073545933, 'clip_ratio/region_mean': 0.05926519073545933, 'reward_total_mean': 0.09551124274730682, 'reward_meter_mean': 0.6407234072685242, 'reward_meter_std': 0.3280641436576843, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.37533053755760193, 'reward_arabic_clean_mean': 0.25, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.7479768395423889, 'reward_lexical_plausibility_std': 0.18183663487434387, 'reward_repeat_penalty_mean': 0.9357457160949707, 'reward_repeat_penalty_std': 0.1146545335650444, 'reward_near_duplicate_penalty_mean': 0.9982457160949707, 'reward_near_duplicate_penalty_std': 0.0018473644740879536, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.16250000894069672, 'reward_judge_quality_std': 0.18077214062213898, 'reward_total_composite_mean': 0.09551124274730682, 'reward_total_composite_std': 0.17697452008724213, 'epoch': 0.02}

 12%|█▏        | 393/3300 [1:28:06<10:11:41, 12.63s/it]INFO 04-12 12:14:18 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 394/3300 [1:28:21<10:37:29, 13.16s/it]2026-04-12 12:14:33,162 | INFO | train_grpo_train | metrics_logged mode=train step=394

                                                       
{'loss': -0.0653, 'grad_norm': 1.706504464149475, 'learning_rate': 8.809090909090909e-06, 'num_tokens': 780953.0, 'completions/mean_length': 277.0, 'completions/min_length': 38.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 42.0, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.7082880735397339, 'rewards/meter/std': 0.41883644461631775, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.3720119297504425, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.8141652941703796, 'rewards/lexical_plausibility/std': 0.20943991839885712, 'rewards/judge_quality/mean': 0.1875, 'rewards/judge_quality/std': 0.16850179433822632, 'rewards/repeat_penalty/mean': 0.9827629923820496, 'rewards/repeat_penalty/std': 0.03876803442835808, 'rewards/near_duplicate_penalty/mean': 0.991126537322998, 'rewards/near_duplicate_penalty/std': 0.016940193250775337, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9912587404251099, 'rewards/distinct_2/std': 0.024724015966057777, 'rewards/total_composite/mean': 0.14978958666324615, 'rewards/total_composite/std': 0.1784851849079132, 'reward': 0.14978958666324615, 'reward_std': 0.1784851849079132, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19560293853282928, 'sampling/sampling_logp_difference/max': 1.159276008605957, 'sampling/importance_sampling_ratio/min': 0.3137132227420807, 'sampling/importance_sampling_ratio/mean': 1.1072908639907837, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4723319709300995, 'clip_ratio/low_mean': 0.01785714365541935, 'clip_ratio/low_min': 0.01785714365541935, 'clip_ratio/high_mean': 0.05713332165032625, 'clip_ratio/high_max': 0.05713332165032625, 'clip_ratio/region_mean': 0.0749904653057456, 'reward_total_mean': 0.14978958666324615, 'reward_meter_mean': 0.7082880735397339, 'reward_meter_std': 0.41883644461631775, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.3720119297504425, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.8141652941703796, 'reward_lexical_plausibility_std': 0.20943991839885712, 'reward_repeat_penalty_mean': 0.9827629923820496, 'reward_repeat_penalty_std': 0.03876803442835808, 'reward_near_duplicate_penalty_mean': 0.991126537322998, 'reward_near_duplicate_penalty_std': 0.016940193250775337, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9912587404251099, 'reward_distinct_2_std': 0.024724015966057777, 'reward_judge_quality_mean': 0.1875, 'reward_judge_quality_std': 0.16850179433822632, 'reward_total_composite_mean': 0.14978958666324615, 'reward_total_composite_std': 0.1784851849079132, 'epoch': 0.02}

 12%|█▏        | 394/3300 [1:28:21<10:37:29, 13.16s/it]INFO 04-12 12:14:33 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 395/3300 [1:28:36<11:10:58, 13.86s/it]2026-04-12 12:14:48,629 | INFO | train_grpo_train | metrics_logged mode=train step=395

                                                       
{'loss': -0.0927, 'grad_norm': 2.346163272857666, 'learning_rate': 8.806060606060608e-06, 'num_tokens': 782715.0, 'completions/mean_length': 292.25, 'completions/min_length': 63.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 72.5, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.2800147533416748, 'rewards/meter/std': 0.36569374799728394, 'rewards/count_adherence/mean': 0.71875, 'rewards/count_adherence/std': 0.41052016615867615, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/lexical_plausibility/mean': 0.8489526510238647, 'rewards/lexical_plausibility/std': 0.16889598965644836, 'rewards/judge_quality/mean': 0.23749999701976776, 'rewards/judge_quality/std': 0.20310096442699432, 'rewards/repeat_penalty/mean': 0.958864152431488, 'rewards/repeat_penalty/std': 0.08593212068080902, 'rewards/near_duplicate_penalty/mean': 0.9951285123825073, 'rewards/near_duplicate_penalty/std': 0.0052598933689296246, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9949392676353455, 'rewards/distinct_2/std': 0.014313912950456142, 'rewards/total_composite/mean': 0.11484209448099136, 'rewards/total_composite/std': 0.17118561267852783, 'reward': 0.11484209448099136, 'reward_std': 0.17118561267852783, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2058657705783844, 'sampling/sampling_logp_difference/max': 0.9940671920776367, 'sampling/importance_sampling_ratio/min': 0.3700684905052185, 'sampling/importance_sampling_ratio/mean': 1.0406041145324707, 'sampling/importance_sampling_ratio/max': 1.8877012729644775, 'entropy': 1.1349950581789017, 'clip_ratio/low_mean': 0.03252242971211672, 'clip_ratio/low_min': 0.03252242971211672, 'clip_ratio/high_mean': 0.04944284074008465, 'clip_ratio/high_max': 0.04944284074008465, 'clip_ratio/region_mean': 0.08196527045220137, 'reward_total_mean': 0.11484209448099136, 'reward_meter_mean': 0.2800147533416748, 'reward_meter_std': 0.36569374799728394, 'reward_count_adherence_mean': 0.71875, 'reward_count_adherence_std': 0.41052016615867615, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_lexical_plausibility_mean': 0.8489526510238647, 'reward_lexical_plausibility_std': 0.16889598965644836, 'reward_repeat_penalty_mean': 0.958864152431488, 'reward_repeat_penalty_std': 0.08593212068080902, 'reward_near_duplicate_penalty_mean': 0.9951285123825073, 'reward_near_duplicate_penalty_std': 0.0052598933689296246, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9949392676353455, 'reward_distinct_2_std': 0.014313912950456142, 'reward_judge_quality_mean': 0.23749999701976776, 'reward_judge_quality_std': 0.20310096442699432, 'reward_total_composite_mean': 0.11484209448099136, 'reward_total_composite_std': 0.17118561267852783, 'epoch': 0.02}

 12%|█▏        | 395/3300 [1:28:36<11:10:58, 13.86s/it]INFO 04-12 12:14:48 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 396/3300 [1:28:51<11:24:37, 14.15s/it]2026-04-12 12:15:03,444 | INFO | train_grpo_train | metrics_logged mode=train step=396

                                                       
{'loss': -0.0659, 'grad_norm': 1.8078728914260864, 'learning_rate': 8.803030303030303e-06, 'num_tokens': 784240.0, 'completions/mean_length': 276.625, 'completions/min_length': 37.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 41.25, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.5899513363838196, 'rewards/meter/std': 0.4457658529281616, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.3720119297504425, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/lexical_plausibility/mean': 0.8389207124710083, 'rewards/lexical_plausibility/std': 0.17937442660331726, 'rewards/judge_quality/mean': 0.3037499785423279, 'rewards/judge_quality/std': 0.31932684779167175, 'rewards/repeat_penalty/mean': 0.9569582939147949, 'rewards/repeat_penalty/std': 0.08557416498661041, 'rewards/near_duplicate_penalty/mean': 0.9882082939147949, 'rewards/near_duplicate_penalty/std': 0.018780311569571495, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2743593156337738, 'rewards/total_composite/std': 0.3373644948005676, 'reward': 0.2743593156337738, 'reward_std': 0.33736446499824524, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19765162467956543, 'sampling/sampling_logp_difference/max': 1.1794836521148682, 'sampling/importance_sampling_ratio/min': 0.3074374496936798, 'sampling/importance_sampling_ratio/mean': 1.073563575744629, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2199176847934723, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.10002384521067142, 'clip_ratio/high_max': 0.10002384521067142, 'clip_ratio/region_mean': 0.10002384521067142, 'reward_total_mean': 0.2743593156337738, 'reward_meter_mean': 0.5899513363838196, 'reward_meter_std': 0.4457658529281616, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.3720119297504425, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_lexical_plausibility_mean': 0.8389207124710083, 'reward_lexical_plausibility_std': 0.17937442660331726, 'reward_repeat_penalty_mean': 0.9569582939147949, 'reward_repeat_penalty_std': 0.08557416498661041, 'reward_near_duplicate_penalty_mean': 0.9882082939147949, 'reward_near_duplicate_penalty_std': 0.018780311569571495, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3037499785423279, 'reward_judge_quality_std': 0.31932684779167175, 'reward_total_composite_mean': 0.2743593156337738, 'reward_total_composite_std': 0.3373644948005676, 'epoch': 0.02}

 12%|█▏        | 396/3300 [1:28:51<11:24:37, 14.15s/it]INFO 04-12 12:15:03 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 397/3300 [1:29:05<11:28:07, 14.22s/it]2026-04-12 12:15:17,846 | INFO | train_grpo_train | metrics_logged mode=train step=397

                                                       
{'loss': -0.1881, 'grad_norm': 3.925553321838379, 'learning_rate': 8.8e-06, 'num_tokens': 786473.0, 'completions/mean_length': 214.125, 'completions/min_length': 95.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 114.83333587646484, 'completions/min_terminated_length': 95.0, 'completions/max_terminated_length': 123.0, 'rewards/meter/mean': 0.6472233533859253, 'rewards/meter/std': 0.41604840755462646, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.3720119297504425, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9404333829879761, 'rewards/lexical_plausibility/std': 0.11923659592866898, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.18077214062213898, 'rewards/repeat_penalty/mean': 0.8763377070426941, 'rewards/repeat_penalty/std': 0.14084605872631073, 'rewards/near_duplicate_penalty/mean': 0.9682947397232056, 'rewards/near_duplicate_penalty/std': 0.026189858093857765, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9296720027923584, 'rewards/distinct_2/std': 0.08126077055931091, 'rewards/total_composite/mean': 0.2649061381816864, 'rewards/total_composite/std': 0.19800522923469543, 'reward': 0.2649061381816864, 'reward_std': 0.19800524413585663, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18909059464931488, 'sampling/sampling_logp_difference/max': 3.081735134124756, 'sampling/importance_sampling_ratio/min': 0.045879580080509186, 'sampling/importance_sampling_ratio/mean': 1.037859320640564, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.005484364926815, 'clip_ratio/low_mean': 0.02631578966975212, 'clip_ratio/low_min': 0.02631578966975212, 'clip_ratio/high_mean': 0.11029411852359772, 'clip_ratio/high_max': 0.11029411852359772, 'clip_ratio/region_mean': 0.13660990819334984, 'reward_total_mean': 0.2649061381816864, 'reward_meter_mean': 0.6472233533859253, 'reward_meter_std': 0.41604840755462646, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.3720119297504425, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9404333829879761, 'reward_lexical_plausibility_std': 0.11923659592866898, 'reward_repeat_penalty_mean': 0.8763377070426941, 'reward_repeat_penalty_std': 0.14084605872631073, 'reward_near_duplicate_penalty_mean': 0.9682947397232056, 'reward_near_duplicate_penalty_std': 0.026189858093857765, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9296720027923584, 'reward_distinct_2_std': 0.08126077055931091, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.18077214062213898, 'reward_total_composite_mean': 0.2649061381816864, 'reward_total_composite_std': 0.19800522923469543, 'epoch': 0.02}

 12%|█▏        | 397/3300 [1:29:05<11:28:07, 14.22s/it]INFO 04-12 12:15:18 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 398/3300 [1:29:20<11:35:01, 14.37s/it]2026-04-12 12:15:32,563 | INFO | train_grpo_train | metrics_logged mode=train step=398

                                                       
{'loss': -0.1388, 'grad_norm': 2.3553307056427, 'learning_rate': 8.796969696969698e-06, 'num_tokens': 788533.0, 'completions/mean_length': 370.5, 'completions/min_length': 129.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.625, 'completions/mean_terminated_length': 134.6666717529297, 'completions/min_terminated_length': 129.0, 'completions/max_terminated_length': 138.0, 'rewards/meter/mean': 0.3942521810531616, 'rewards/meter/std': 0.36259716749191284, 'rewards/count_adherence/mean': 0.5535714626312256, 'rewards/count_adherence/std': 0.36967799067497253, 'rewards/arabic_clean/mean': 0.375, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.8397576212882996, 'rewards/lexical_plausibility/std': 0.1545770764350891, 'rewards/judge_quality/mean': 0.17500001192092896, 'rewards/judge_quality/std': 0.1752549111843109, 'rewards/repeat_penalty/mean': 0.9140479564666748, 'rewards/repeat_penalty/std': 0.2199569195508957, 'rewards/near_duplicate_penalty/mean': 0.9818993806838989, 'rewards/near_duplicate_penalty/std': 0.03753954544663429, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9321063756942749, 'rewards/distinct_2/std': 0.18265652656555176, 'rewards/total_composite/mean': 0.10109606385231018, 'rewards/total_composite/std': 0.1619570106267929, 'reward': 0.10109606385231018, 'reward_std': 0.1619569957256317, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18112711608409882, 'sampling/sampling_logp_difference/max': 1.2439486980438232, 'sampling/importance_sampling_ratio/min': 0.2882438004016876, 'sampling/importance_sampling_ratio/mean': 1.0341545343399048, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6858463287353516, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.065457409247756, 'clip_ratio/high_max': 0.065457409247756, 'clip_ratio/region_mean': 0.065457409247756, 'reward_total_mean': 0.10109606385231018, 'reward_meter_mean': 0.3942521810531616, 'reward_meter_std': 0.36259716749191284, 'reward_count_adherence_mean': 0.5535714626312256, 'reward_count_adherence_std': 0.36967799067497253, 'reward_arabic_clean_mean': 0.375, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.8397576212882996, 'reward_lexical_plausibility_std': 0.1545770764350891, 'reward_repeat_penalty_mean': 0.9140479564666748, 'reward_repeat_penalty_std': 0.2199569195508957, 'reward_near_duplicate_penalty_mean': 0.9818993806838989, 'reward_near_duplicate_penalty_std': 0.03753954544663429, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9321063756942749, 'reward_distinct_2_std': 0.18265652656555176, 'reward_judge_quality_mean': 0.17500001192092896, 'reward_judge_quality_std': 0.1752549111843109, 'reward_total_composite_mean': 0.10109606385231018, 'reward_total_composite_std': 0.1619570106267929, 'epoch': 0.02}

 12%|█▏        | 398/3300 [1:29:20<11:35:01, 14.37s/it]INFO 04-12 12:15:32 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 399/3300 [1:29:29<10:19:24, 12.81s/it]2026-04-12 12:15:41,735 | INFO | train_grpo_train | metrics_logged mode=train step=399

                                                       
{'loss': 0.0055, 'grad_norm': 15.326539039611816, 'learning_rate': 8.793939393939395e-06, 'num_tokens': 790239.0, 'completions/mean_length': 41.25, 'completions/min_length': 36.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.25, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.9389407634735107, 'rewards/meter/std': 0.11409443616867065, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.71875, 'rewards/judge_quality/std': 0.2794605791568756, 'rewards/repeat_penalty/mean': 0.9888359308242798, 'rewards/repeat_penalty/std': 0.014076457358896732, 'rewards/near_duplicate_penalty/mean': 0.9888359308242798, 'rewards/near_duplicate_penalty/std': 0.014076457358896732, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6672719717025757, 'rewards/total_composite/std': 0.26418986916542053, 'reward': 0.6672719717025757, 'reward_std': 0.26418986916542053, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1759723275899887, 'sampling/sampling_logp_difference/max': 2.0261192321777344, 'sampling/importance_sampling_ratio/min': 0.13184620440006256, 'sampling/importance_sampling_ratio/mean': 1.0538458824157715, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5416480898857117, 'clip_ratio/low_mean': 0.06201437674462795, 'clip_ratio/low_min': 0.06201437674462795, 'clip_ratio/high_mean': 0.08641835395246744, 'clip_ratio/high_max': 0.08641835395246744, 'clip_ratio/region_mean': 0.1484327306970954, 'reward_total_mean': 0.6672719717025757, 'reward_meter_mean': 0.9389407634735107, 'reward_meter_std': 0.11409443616867065, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9888359308242798, 'reward_repeat_penalty_std': 0.014076457358896732, 'reward_near_duplicate_penalty_mean': 0.9888359308242798, 'reward_near_duplicate_penalty_std': 0.014076457358896732, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.71875, 'reward_judge_quality_std': 0.2794605791568756, 'reward_total_composite_mean': 0.6672719717025757, 'reward_total_composite_std': 0.26418986916542053, 'epoch': 0.02}

 12%|█▏        | 399/3300 [1:29:29<10:19:24, 12.81s/it]INFO 04-12 12:15:41 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 400/3300 [1:29:39<9:35:23, 11.90s/it] 2026-04-12 12:15:51,524 | INFO | train_grpo_train | metrics_logged mode=train step=400

                                                      
{'loss': -0.0154, 'grad_norm': 12.646838188171387, 'learning_rate': 8.790909090909092e-06, 'num_tokens': 792163.0, 'completions/mean_length': 63.5, 'completions/min_length': 51.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.5, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.8272651433944702, 'rewards/meter/std': 0.3097221553325653, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.48749998211860657, 'rewards/judge_quality/std': 0.2711220383644104, 'rewards/repeat_penalty/mean': 0.9963112473487854, 'rewards/repeat_penalty/std': 0.004177942406386137, 'rewards/near_duplicate_penalty/mean': 0.9963112473487854, 'rewards/near_duplicate_penalty/std': 0.004177942406386137, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4300561845302582, 'rewards/total_composite/std': 0.31272780895233154, 'reward': 0.4300561845302582, 'reward_std': 0.31272780895233154, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18742048740386963, 'sampling/sampling_logp_difference/max': 1.7906854152679443, 'sampling/importance_sampling_ratio/min': 0.16684576869010925, 'sampling/importance_sampling_ratio/mean': 1.019217848777771, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6776831150054932, 'clip_ratio/low_mean': 0.13417689688503742, 'clip_ratio/low_min': 0.13417689688503742, 'clip_ratio/high_mean': 0.026569634675979614, 'clip_ratio/high_max': 0.026569634675979614, 'clip_ratio/region_mean': 0.16074653156101704, 'reward_total_mean': 0.4300561845302582, 'reward_meter_mean': 0.8272651433944702, 'reward_meter_std': 0.3097221553325653, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9963112473487854, 'reward_repeat_penalty_std': 0.004177942406386137, 'reward_near_duplicate_penalty_mean': 0.9963112473487854, 'reward_near_duplicate_penalty_std': 0.004177942406386137, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.48749998211860657, 'reward_judge_quality_std': 0.2711220383644104, 'reward_total_composite_mean': 0.4300561845302582, 'reward_total_composite_std': 0.31272780895233154, 'epoch': 0.02}

 12%|█▏        | 400/3300 [1:29:39<9:35:23, 11.90s/it]INFO 04-12 12:15:51 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


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

 15%|█▌        | 2/13 [00:12<01:06,  6.01s/it]

 23%|██▎       | 3/13 [00:20<01:09,  6.95s/it]

 31%|███       | 4/13 [00:32<01:19,  8.87s/it]

 38%|███▊      | 5/13 [00:44<01:19, 10.00s/it]

 46%|████▌     | 6/13 [00:56<01:15, 10.80s/it]

 54%|█████▍    | 7/13 [01:09<01:07, 11.25s/it]

 62%|██████▏   | 8/13 [01:22<00:59, 11.80s/it]

 69%|██████▉   | 9/13 [01:34<00:47, 11.89s/it]

 77%|███████▋  | 10/13 [01:46<00:36, 12.08s/it]

 85%|████████▍ | 11/13 [01:58<00:23, 11.85s/it]

 92%|█████████▏| 12/13 [02:10<00:11, 11.96s/it]

100%|██████████| 13/13 [02:22<00:00, 11.96s/it]2026-04-12 12:18:25,773 | INFO | train_grpo_train | metrics_logged mode=eval step=400

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 154.2168, 'eval_samples_per_second': 0.674, 'eval_steps_per_second': 0.084, 'eval_num_tokens': 792163.0, 'eval_completions/mean_length': 187.70192307692307, 'eval_completions/min_length': 38.23076923076923, 'eval_completions/max_length': 471.9230769230769, 'eval_completions/clipped_ratio': 0.14423076923076922, 'eval_completions/mean_terminated_length': 132.26016822228064, 'eval_completions/min_terminated_length': 38.23076923076923, 'eval_completions/max_terminated_length': 286.15384615384613, 'eval_rewards/meter/mean': 0.5420474341282477, 'eval_rewards/meter/std': 0.3766455558630136, 'eval_rewards/count_adherence/mean': 0.8512307313772348, 'eval_rewards/count_adherence/std': 0.2499035453567138, 'eval_rewards/arabic_clean/mean': 0.7788461538461539, 'eval_rewards/arabic_clean/std': 0.38037164165423465, 'eval_rewards/lexical_plausibility/mean': 0.9559404253959656, 'eval_rewards/lexical_plausibility/std': 0.08785561610084887, 'eval_rewards/judge_quality/mean': 0.3762499965154208, 'eval_rewards/judge_quality/std': 0.2364156165948281, 'eval_rewards/repeat_penalty/mean': 0.8230979901093703, 'eval_rewards/repeat_penalty/std': 0.2612293680699972, 'eval_rewards/near_duplicate_penalty/mean': 0.9467151898604172, 'eval_rewards/near_duplicate_penalty/std': 0.08159351735734023, 'eval_rewards/opening_diversity/mean': 0.9554265416585482, 'eval_rewards/opening_diversity/std': 0.0883646491341866, 'eval_rewards/distinct_2/mean': 0.8715694730098431, 'eval_rewards/distinct_2/std': 0.226575913767402, 'eval_rewards/total_composite/mean': 0.190441376887835, 'eval_rewards/total_composite/std': 0.17775888225206962, 'eval_reward': 0.190441376887835, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.10218952195002483, 'eval_sampling/sampling_logp_difference/max': 1.2410474190345178, 'eval_sampling/importance_sampling_ratio/min': 0.2915732367680623, 'eval_sampling/importance_sampling_ratio/mean': 1.0252403112558217, 'eval_sampling/importance_sampling_ratio/max': 1.5982536994493926, 'eval_entropy': 1.3140468368163476, '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': 0.190441376887835, 'eval_reward_meter_mean': 0.5420474341282477, 'eval_reward_meter_std': 0.3766455558630136, 'eval_reward_count_adherence_mean': 0.8512307313772348, 'eval_reward_count_adherence_std': 0.2499035453567138, 'eval_reward_arabic_clean_mean': 0.7788461538461539, 'eval_reward_arabic_clean_std': 0.38037164165423465, 'eval_reward_lexical_plausibility_mean': 0.9559404253959656, 'eval_reward_lexical_plausibility_std': 0.08785561610084887, 'eval_reward_repeat_penalty_mean': 0.8230979901093703, 'eval_reward_repeat_penalty_std': 0.2612293680699972, 'eval_reward_near_duplicate_penalty_mean': 0.9467151898604172, 'eval_reward_near_duplicate_penalty_std': 0.08159351735734023, 'eval_reward_opening_diversity_mean': 0.9554265416585482, 'eval_reward_opening_diversity_std': 0.0883646491341866, 'eval_reward_distinct_2_mean': 0.8715694730098431, 'eval_reward_distinct_2_std': 0.226575913767402, 'eval_reward_judge_quality_mean': 0.3762499965154208, 'eval_reward_judge_quality_std': 0.2364156165948281, 'eval_reward_total_composite_mean': 0.190441376887835, 'eval_reward_total_composite_std': 0.17775888225206962, 'epoch': 0.02}

 12%|█▏        | 400/3300 [1:32:13<9:35:23, 11.90s/it]

100%|██████████| 13/13 [02:22<00:00, 11.96s/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(

 12%|█▏        | 401/3300 [1:32:25<46:52:05, 58.20s/it]2026-04-12 12:18:37,767 | INFO | train_grpo_train | metrics_logged mode=train step=401

                                                       
{'loss': -0.0296, 'grad_norm': 16.03925895690918, 'learning_rate': 8.787878787878788e-06, 'num_tokens': 793759.0, 'completions/mean_length': 42.5, 'completions/min_length': 38.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.5, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.8769299983978271, 'rewards/meter/std': 0.30164098739624023, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7512500286102295, 'rewards/judge_quality/std': 0.24976776540279388, 'rewards/repeat_penalty/mean': 0.9924679398536682, 'rewards/repeat_penalty/std': 0.00984637625515461, 'rewards/near_duplicate_penalty/mean': 0.9924679398536682, 'rewards/near_duplicate_penalty/std': 0.00984637625515461, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6401259899139404, 'rewards/total_composite/std': 0.3147824704647064, 'reward': 0.6401259899139404, 'reward_std': 0.3147824704647064, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1906818002462387, 'sampling/sampling_logp_difference/max': 2.0528461933135986, 'sampling/importance_sampling_ratio/min': 0.1283690184354782, 'sampling/importance_sampling_ratio/mean': 1.0127277374267578, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4589977413415909, 'clip_ratio/low_mean': 0.06584052927792072, 'clip_ratio/low_min': 0.06584052927792072, 'clip_ratio/high_mean': 0.08660051599144936, 'clip_ratio/high_max': 0.08660051599144936, 'clip_ratio/region_mean': 0.15244104526937008, 'reward_total_mean': 0.6401259899139404, 'reward_meter_mean': 0.8769299983978271, 'reward_meter_std': 0.30164098739624023, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9924679398536682, 'reward_repeat_penalty_std': 0.00984637625515461, 'reward_near_duplicate_penalty_mean': 0.9924679398536682, 'reward_near_duplicate_penalty_std': 0.00984637625515461, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7512500286102295, 'reward_judge_quality_std': 0.24976776540279388, 'reward_total_composite_mean': 0.6401259899139404, 'reward_total_composite_std': 0.3147824704647064, 'epoch': 0.02}

 12%|█▏        | 401/3300 [1:32:25<46:52:05, 58.20s/it]INFO 04-12 12:18:38 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 402/3300 [1:32:40<36:21:40, 45.17s/it]2026-04-12 12:18:52,523 | INFO | train_grpo_train | metrics_logged mode=train step=402

                                                       
{'loss': -0.0929, 'grad_norm': 5.5130510330200195, 'learning_rate': 8.784848484848487e-06, 'num_tokens': 795630.0, 'completions/mean_length': 125.875, 'completions/min_length': 52.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 70.71428680419922, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 88.0, 'rewards/meter/mean': 0.7937238216400146, 'rewards/meter/std': 0.3507080078125, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.3471825420856476, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9805421829223633, 'rewards/lexical_plausibility/std': 0.049931079149246216, 'rewards/judge_quality/mean': 0.4050000309944153, 'rewards/judge_quality/std': 0.33713075518608093, 'rewards/repeat_penalty/mean': 0.9519250988960266, 'rewards/repeat_penalty/std': 0.07706359773874283, 'rewards/near_duplicate_penalty/mean': 0.9747341871261597, 'rewards/near_duplicate_penalty/std': 0.03452078998088837, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9751819372177124, 'rewards/distinct_2/std': 0.04675424471497536, 'rewards/total_composite/mean': 0.319842666387558, 'rewards/total_composite/std': 0.2840674817562103, 'reward': 0.319842666387558, 'reward_std': 0.2840674817562103, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19560658931732178, 'sampling/sampling_logp_difference/max': 1.263073205947876, 'sampling/importance_sampling_ratio/min': 0.2827836275100708, 'sampling/importance_sampling_ratio/mean': 1.041720986366272, 'sampling/importance_sampling_ratio/max': 1.9923841953277588, 'entropy': 1.898254171013832, 'clip_ratio/low_mean': 0.10888747870922089, 'clip_ratio/low_min': 0.10888747870922089, 'clip_ratio/high_mean': 0.07550730556249619, 'clip_ratio/high_max': 0.07550730556249619, 'clip_ratio/region_mean': 0.18439478427171707, 'reward_total_mean': 0.319842666387558, 'reward_meter_mean': 0.7937238216400146, 'reward_meter_std': 0.3507080078125, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.3471825420856476, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9805421829223633, 'reward_lexical_plausibility_std': 0.049931079149246216, 'reward_repeat_penalty_mean': 0.9519250988960266, 'reward_repeat_penalty_std': 0.07706359773874283, 'reward_near_duplicate_penalty_mean': 0.9747341871261597, 'reward_near_duplicate_penalty_std': 0.03452078998088837, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9751819372177124, 'reward_distinct_2_std': 0.04675424471497536, 'reward_judge_quality_mean': 0.4050000309944153, 'reward_judge_quality_std': 0.33713075518608093, 'reward_total_composite_mean': 0.319842666387558, 'reward_total_composite_std': 0.2840674817562103, 'epoch': 0.02}

 12%|█▏        | 402/3300 [1:32:40<36:21:40, 45.17s/it]INFO 04-12 12:18:52 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 403/3300 [1:32:56<29:17:09, 36.39s/it]2026-04-12 12:19:08,435 | INFO | train_grpo_train | metrics_logged mode=train step=403

                                                       
{'loss': -0.0867, 'grad_norm': 5.634540557861328, 'learning_rate': 8.781818181818182e-06, 'num_tokens': 797499.0, 'completions/mean_length': 139.625, 'completions/min_length': 75.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 86.42857360839844, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 116.0, 'rewards/meter/mean': 0.7045705914497375, 'rewards/meter/std': 0.36760589480400085, '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/lexical_plausibility/mean': 0.959592878818512, 'rewards/lexical_plausibility/std': 0.09168154746294022, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.11649646610021591, 'rewards/repeat_penalty/mean': 0.8338491916656494, 'rewards/repeat_penalty/std': 0.1686556488275528, 'rewards/near_duplicate_penalty/mean': 0.9704387784004211, 'rewards/near_duplicate_penalty/std': 0.020224422216415405, 'rewards/opening_diversity/mean': 0.9453125, 'rewards/opening_diversity/std': 0.08476267009973526, 'rewards/distinct_2/mean': 0.9160425662994385, 'rewards/distinct_2/std': 0.08388027548789978, 'rewards/total_composite/mean': 0.25078490376472473, 'rewards/total_composite/std': 0.1559588611125946, 'reward': 0.25078490376472473, 'reward_std': 0.1559588611125946, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16822002828121185, 'sampling/sampling_logp_difference/max': 1.9854364395141602, 'sampling/importance_sampling_ratio/min': 0.13732066750526428, 'sampling/importance_sampling_ratio/mean': 1.022871971130371, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2079317420721054, 'clip_ratio/low_mean': 0.04398762248456478, 'clip_ratio/low_min': 0.04398762248456478, 'clip_ratio/high_mean': 0.1001646718941629, 'clip_ratio/high_max': 0.1001646718941629, 'clip_ratio/region_mean': 0.14415229437872767, 'reward_total_mean': 0.25078490376472473, 'reward_meter_mean': 0.7045705914497375, 'reward_meter_std': 0.36760589480400085, '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_lexical_plausibility_mean': 0.959592878818512, 'reward_lexical_plausibility_std': 0.09168154746294022, 'reward_repeat_penalty_mean': 0.8338491916656494, 'reward_repeat_penalty_std': 0.1686556488275528, 'reward_near_duplicate_penalty_mean': 0.9704387784004211, 'reward_near_duplicate_penalty_std': 0.020224422216415405, 'reward_opening_diversity_mean': 0.9453125, 'reward_opening_diversity_std': 0.08476267009973526, 'reward_distinct_2_mean': 0.9160425662994385, 'reward_distinct_2_std': 0.08388027548789978, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.11649646610021591, 'reward_total_composite_mean': 0.25078490376472473, 'reward_total_composite_std': 0.1559588611125946, 'epoch': 0.02}

 12%|█▏        | 403/3300 [1:32:56<29:17:09, 36.39s/it]INFO 04-12 12:19:08 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 404/3300 [1:33:12<24:21:07, 30.27s/it]2026-04-12 12:19:24,418 | INFO | train_grpo_train | metrics_logged mode=train step=404

                                                       
{'loss': -0.0718, 'grad_norm': 1.6249110698699951, 'learning_rate': 8.77878787878788e-06, 'num_tokens': 799057.0, 'completions/mean_length': 279.75, 'completions/min_length': 42.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 47.5, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.9419029355049133, 'rewards/meter/std': 0.1101876050233841, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/lexical_plausibility/mean': 0.8911904096603394, 'rewards/lexical_plausibility/std': 0.10423477739095688, 'rewards/judge_quality/mean': 0.3087500035762787, 'rewards/judge_quality/std': 0.3165184557437897, 'rewards/repeat_penalty/mean': 0.9617950916290283, 'rewards/repeat_penalty/std': 0.08661969751119614, 'rewards/near_duplicate_penalty/mean': 0.9935762882232666, 'rewards/near_duplicate_penalty/std': 0.013885052874684334, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9994688034057617, 'rewards/distinct_2/std': 0.00150255614425987, 'rewards/total_composite/mean': 0.27821362018585205, 'rewards/total_composite/std': 0.3338005840778351, 'reward': 0.27821362018585205, 'reward_std': 0.3338005840778351, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19768500328063965, 'sampling/sampling_logp_difference/max': 1.7469000816345215, 'sampling/importance_sampling_ratio/min': 0.17431345582008362, 'sampling/importance_sampling_ratio/mean': 1.0551173686981201, 'sampling/importance_sampling_ratio/max': 1.8611770868301392, 'entropy': 1.2844100147485733, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.07779906783252954, 'clip_ratio/high_max': 0.07779906783252954, 'clip_ratio/region_mean': 0.07779906783252954, 'reward_total_mean': 0.27821362018585205, 'reward_meter_mean': 0.9419029355049133, 'reward_meter_std': 0.1101876050233841, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_lexical_plausibility_mean': 0.8911904096603394, 'reward_lexical_plausibility_std': 0.10423477739095688, 'reward_repeat_penalty_mean': 0.9617950916290283, 'reward_repeat_penalty_std': 0.08661969751119614, 'reward_near_duplicate_penalty_mean': 0.9935762882232666, 'reward_near_duplicate_penalty_std': 0.013885052874684334, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9994688034057617, 'reward_distinct_2_std': 0.00150255614425987, 'reward_judge_quality_mean': 0.3087500035762787, 'reward_judge_quality_std': 0.3165184557437897, 'reward_total_composite_mean': 0.27821362018585205, 'reward_total_composite_std': 0.3338005840778351, 'epoch': 0.02}

 12%|█▏        | 404/3300 [1:33:12<24:21:07, 30.27s/it]INFO 04-12 12:19:24 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 405/3300 [1:33:21<19:20:42, 24.06s/it]2026-04-12 12:19:33,983 | INFO | train_grpo_train | metrics_logged mode=train step=405

                                                       
{'loss': 0.0602, 'grad_norm': 21.173933029174805, 'learning_rate': 8.775757575757577e-06, 'num_tokens': 800562.0, 'completions/mean_length': 21.125, 'completions/min_length': 18.0, 'completions/max_length': 24.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.125, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.9415585994720459, 'rewards/meter/std': 0.09394428133964539, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.9387500286102295, 'rewards/judge_quality/std': 0.015526460483670235, 'rewards/repeat_penalty/mean': 0.7492251992225647, 'rewards/repeat_penalty/std': 0.002191446488723159, 'rewards/near_duplicate_penalty/mean': 0.9989669322967529, 'rewards/near_duplicate_penalty/std': 0.002921935636550188, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.883629322052002, 'rewards/total_composite/std': 0.087247833609581, 'reward': 0.883629322052002, 'reward_std': 0.0872478261590004, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13938158750534058, 'sampling/sampling_logp_difference/max': 1.0163929462432861, 'sampling/importance_sampling_ratio/min': 0.3618979752063751, 'sampling/importance_sampling_ratio/mean': 1.0382153987884521, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9832390174269676, 'clip_ratio/low_mean': 0.060132576152682304, 'clip_ratio/low_min': 0.060132576152682304, 'clip_ratio/high_mean': 0.10510136000812054, 'clip_ratio/high_max': 0.10510136000812054, 'clip_ratio/region_mean': 0.16523393616080284, 'reward_total_mean': 0.883629322052002, 'reward_meter_mean': 0.9415585994720459, 'reward_meter_std': 0.09394428133964539, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7492251992225647, 'reward_repeat_penalty_std': 0.002191446488723159, 'reward_near_duplicate_penalty_mean': 0.9989669322967529, 'reward_near_duplicate_penalty_std': 0.002921935636550188, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.9387500286102295, 'reward_judge_quality_std': 0.015526460483670235, 'reward_total_composite_mean': 0.883629322052002, 'reward_total_composite_std': 0.087247833609581, 'epoch': 0.02}

 12%|█▏        | 405/3300 [1:33:21<19:20:42, 24.06s/it]INFO 04-12 12:19:34 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 406/3300 [1:33:30<15:42:31, 19.54s/it]2026-04-12 12:19:42,983 | INFO | train_grpo_train | metrics_logged mode=train step=406

                                                       
{'loss': 0.057, 'grad_norm': 22.811269760131836, 'learning_rate': 8.772727272727274e-06, 'num_tokens': 801864.0, 'completions/mean_length': 21.75, 'completions/min_length': 20.0, 'completions/max_length': 25.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.75, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.7749293446540833, 'rewards/meter/std': 0.3158000111579895, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9791666269302368, 'rewards/lexical_plausibility/std': 0.0589255727827549, 'rewards/judge_quality/mean': 0.8612500429153442, 'rewards/judge_quality/std': 0.16617010533809662, 'rewards/repeat_penalty/mean': 0.7400568127632141, 'rewards/repeat_penalty/std': 0.02386556752026081, 'rewards/near_duplicate_penalty/mean': 0.9867424368858337, 'rewards/near_duplicate_penalty/std': 0.03182074800133705, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6544976234436035, 'rewards/total_composite/std': 0.2910785675048828, 'reward': 0.6544976234436035, 'reward_std': 0.2910785377025604, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14497463405132294, 'sampling/sampling_logp_difference/max': 1.519446849822998, 'sampling/importance_sampling_ratio/min': 0.21883292496204376, 'sampling/importance_sampling_ratio/mean': 0.9946885108947754, 'sampling/importance_sampling_ratio/max': 1.9460806846618652, 'entropy': 0.8249792009592056, 'clip_ratio/low_mean': 0.05614389292895794, 'clip_ratio/low_min': 0.05614389292895794, 'clip_ratio/high_mean': 0.0941381985321641, 'clip_ratio/high_max': 0.0941381985321641, 'clip_ratio/region_mean': 0.15028209146112204, 'reward_total_mean': 0.6544976234436035, 'reward_meter_mean': 0.7749293446540833, 'reward_meter_std': 0.3158000111579895, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9791666269302368, 'reward_lexical_plausibility_std': 0.0589255727827549, 'reward_repeat_penalty_mean': 0.7400568127632141, 'reward_repeat_penalty_std': 0.02386556752026081, 'reward_near_duplicate_penalty_mean': 0.9867424368858337, 'reward_near_duplicate_penalty_std': 0.03182074800133705, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8612500429153442, 'reward_judge_quality_std': 0.16617010533809662, 'reward_total_composite_mean': 0.6544976234436035, 'reward_total_composite_std': 0.2910785675048828, 'epoch': 0.02}

 12%|█▏        | 406/3300 [1:33:30<15:42:31, 19.54s/it]INFO 04-12 12:19:43 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 407/3300 [1:33:39<13:05:23, 16.29s/it]2026-04-12 12:19:51,682 | INFO | train_grpo_train | metrics_logged mode=train step=407

                                                       
{'loss': 0.0167, 'grad_norm': 15.969156265258789, 'learning_rate': 8.76969696969697e-06, 'num_tokens': 803547.0, 'completions/mean_length': 40.375, 'completions/min_length': 32.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.375, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.5834869146347046, 'rewards/meter/std': 0.38459333777427673, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9866071343421936, 'rewards/lexical_plausibility/std': 0.03788072615861893, 'rewards/judge_quality/mean': 0.7012500166893005, 'rewards/judge_quality/std': 0.21963852643966675, 'rewards/repeat_penalty/mean': 0.9848974943161011, 'rewards/repeat_penalty/std': 0.01567939855158329, 'rewards/near_duplicate_penalty/mean': 0.9848974943161011, 'rewards/near_duplicate_penalty/std': 0.01567939855158329, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.349627822637558, 'rewards/total_composite/std': 0.17711110413074493, 'reward': 0.349627822637558, 'reward_std': 0.17711110413074493, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18548497557640076, 'sampling/sampling_logp_difference/max': 1.9679288864135742, 'sampling/importance_sampling_ratio/min': 0.13974598050117493, 'sampling/importance_sampling_ratio/mean': 1.0032404661178589, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9753029942512512, 'clip_ratio/low_mean': 0.05670153442770243, 'clip_ratio/low_min': 0.05670153442770243, 'clip_ratio/high_mean': 0.10682980204001069, 'clip_ratio/high_max': 0.10682980204001069, 'clip_ratio/region_mean': 0.16353133646771312, 'reward_total_mean': 0.349627822637558, 'reward_meter_mean': 0.5834869146347046, 'reward_meter_std': 0.38459333777427673, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9866071343421936, 'reward_lexical_plausibility_std': 0.03788072615861893, 'reward_repeat_penalty_mean': 0.9848974943161011, 'reward_repeat_penalty_std': 0.01567939855158329, 'reward_near_duplicate_penalty_mean': 0.9848974943161011, 'reward_near_duplicate_penalty_std': 0.01567939855158329, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7012500166893005, 'reward_judge_quality_std': 0.21963852643966675, 'reward_total_composite_mean': 0.349627822637558, 'reward_total_composite_std': 0.17711110413074493, 'epoch': 0.02}

 12%|█▏        | 407/3300 [1:33:39<13:05:23, 16.29s/it]INFO 04-12 12:19:51 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 408/3300 [1:33:49<11:38:01, 14.48s/it]2026-04-12 12:20:01,944 | INFO | train_grpo_train | metrics_logged mode=train step=408

                                                       
{'loss': 0.0098, 'grad_norm': 23.339908599853516, 'learning_rate': 8.766666666666669e-06, 'num_tokens': 805135.0, 'completions/mean_length': 39.5, 'completions/min_length': 31.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.5, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.745948314666748, 'rewards/meter/std': 0.36162394285202026, '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/lexical_plausibility/mean': 0.9822304248809814, 'rewards/lexical_plausibility/std': 0.0502600334584713, 'rewards/judge_quality/mean': 0.4424999952316284, 'rewards/judge_quality/std': 0.32101404666900635, 'rewards/repeat_penalty/mean': 0.9928247928619385, 'rewards/repeat_penalty/std': 0.009688506834208965, 'rewards/near_duplicate_penalty/mean': 0.9928247928619385, 'rewards/near_duplicate_penalty/std': 0.009688506834208965, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3352487087249756, 'rewards/total_composite/std': 0.38816389441490173, 'reward': 0.3352487087249756, 'reward_std': 0.38816389441490173, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1757953017950058, 'sampling/sampling_logp_difference/max': 2.326653242111206, 'sampling/importance_sampling_ratio/min': 0.09762191027402878, 'sampling/importance_sampling_ratio/mean': 1.0229308605194092, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1553932651877403, 'clip_ratio/low_mean': 0.08606366440653801, 'clip_ratio/low_min': 0.08606366440653801, 'clip_ratio/high_mean': 0.0677096527069807, 'clip_ratio/high_max': 0.0677096527069807, 'clip_ratio/region_mean': 0.15377331711351871, 'reward_total_mean': 0.3352487087249756, 'reward_meter_mean': 0.745948314666748, 'reward_meter_std': 0.36162394285202026, '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_lexical_plausibility_mean': 0.9822304248809814, 'reward_lexical_plausibility_std': 0.0502600334584713, 'reward_repeat_penalty_mean': 0.9928247928619385, 'reward_repeat_penalty_std': 0.009688506834208965, 'reward_near_duplicate_penalty_mean': 0.9928247928619385, 'reward_near_duplicate_penalty_std': 0.009688506834208965, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4424999952316284, 'reward_judge_quality_std': 0.32101404666900635, 'reward_total_composite_mean': 0.3352487087249756, 'reward_total_composite_std': 0.38816389441490173, 'epoch': 0.02}

 12%|█▏        | 408/3300 [1:33:49<11:38:01, 14.48s/it]INFO 04-12 12:20:02 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 409/3300 [1:33:58<10:12:06, 12.70s/it]2026-04-12 12:20:10,497 | INFO | train_grpo_train | metrics_logged mode=train step=409

                                                       
{'loss': 0.1073, 'grad_norm': 12.45975399017334, 'learning_rate': 8.763636363636364e-06, 'num_tokens': 806699.0, 'completions/mean_length': 44.5, 'completions/min_length': 35.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.5, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.8550735712051392, 'rewards/meter/std': 0.2305542677640915, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6262499690055847, 'rewards/judge_quality/std': 0.3325631022453308, 'rewards/repeat_penalty/mean': 0.9285362958908081, 'rewards/repeat_penalty/std': 0.09516895562410355, 'rewards/near_duplicate_penalty/mean': 0.9433305263519287, 'rewards/near_duplicate_penalty/std': 0.06342414021492004, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9825174808502197, 'rewards/distinct_2/std': 0.049448031932115555, 'rewards/total_composite/mean': 0.492623895406723, 'rewards/total_composite/std': 0.3794953525066376, 'reward': 0.492623895406723, 'reward_std': 0.3794953525066376, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16851359605789185, 'sampling/sampling_logp_difference/max': 1.8270151615142822, 'sampling/importance_sampling_ratio/min': 0.16089309751987457, 'sampling/importance_sampling_ratio/mean': 1.0132133960723877, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4425849094986916, 'clip_ratio/low_mean': 0.09101798571646214, 'clip_ratio/low_min': 0.09101798571646214, 'clip_ratio/high_mean': 0.05231028236448765, 'clip_ratio/high_max': 0.05231028236448765, 'clip_ratio/region_mean': 0.14332826808094978, 'reward_total_mean': 0.492623895406723, 'reward_meter_mean': 0.8550735712051392, 'reward_meter_std': 0.2305542677640915, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9285362958908081, 'reward_repeat_penalty_std': 0.09516895562410355, 'reward_near_duplicate_penalty_mean': 0.9433305263519287, 'reward_near_duplicate_penalty_std': 0.06342414021492004, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9825174808502197, 'reward_distinct_2_std': 0.049448031932115555, 'reward_judge_quality_mean': 0.6262499690055847, 'reward_judge_quality_std': 0.3325631022453308, 'reward_total_composite_mean': 0.492623895406723, 'reward_total_composite_std': 0.3794953525066376, 'epoch': 0.02}

 12%|█▏        | 409/3300 [1:33:58<10:12:06, 12.70s/it]INFO 04-12 12:20:10 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 410/3300 [1:34:08<9:28:16, 11.80s/it] 2026-04-12 12:20:20,185 | INFO | train_grpo_train | metrics_logged mode=train step=410

                                                      
{'loss': 0.0009, 'grad_norm': 12.330366134643555, 'learning_rate': 8.760606060606061e-06, 'num_tokens': 808932.0, 'completions/mean_length': 87.125, 'completions/min_length': 80.0, 'completions/max_length': 95.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 87.125, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 95.0, 'rewards/meter/mean': 0.6178456544876099, 'rewards/meter/std': 0.32561543583869934, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.41249996423721313, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.9310047626495361, 'rewards/repeat_penalty/std': 0.07493745535612106, 'rewards/near_duplicate_penalty/mean': 0.9718864560127258, 'rewards/near_duplicate_penalty/std': 0.022444119676947594, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9568679928779602, 'rewards/distinct_2/std': 0.05820569396018982, 'rewards/total_composite/mean': 0.2374134659767151, 'rewards/total_composite/std': 0.1463068127632141, 'reward': 0.2374134659767151, 'reward_std': 0.1463068127632141, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1925693303346634, 'sampling/sampling_logp_difference/max': 3.0074214935302734, 'sampling/importance_sampling_ratio/min': 0.049418944865465164, 'sampling/importance_sampling_ratio/mean': 1.0061984062194824, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5813437700271606, 'clip_ratio/low_mean': 0.07369900029152632, 'clip_ratio/low_min': 0.07369900029152632, 'clip_ratio/high_mean': 0.08327173069119453, 'clip_ratio/high_max': 0.08327173069119453, 'clip_ratio/region_mean': 0.15697073098272085, 'reward_total_mean': 0.2374134659767151, 'reward_meter_mean': 0.6178456544876099, 'reward_meter_std': 0.32561543583869934, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9310047626495361, 'reward_repeat_penalty_std': 0.07493745535612106, 'reward_near_duplicate_penalty_mean': 0.9718864560127258, 'reward_near_duplicate_penalty_std': 0.022444119676947594, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9568679928779602, 'reward_distinct_2_std': 0.05820569396018982, 'reward_judge_quality_mean': 0.41249996423721313, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.2374134659767151, 'reward_total_composite_std': 0.1463068127632141, 'epoch': 0.02}

 12%|█▏        | 410/3300 [1:34:08<9:28:16, 11.80s/it]INFO 04-12 12:20:20 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 411/3300 [1:34:17<8:54:12, 11.09s/it]2026-04-12 12:20:29,640 | INFO | train_grpo_train | metrics_logged mode=train step=411

                                                      
{'loss': 0.0698, 'grad_norm': 19.95772361755371, 'learning_rate': 8.757575757575759e-06, 'num_tokens': 810595.0, 'completions/mean_length': 39.875, 'completions/min_length': 36.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.875, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.6561762094497681, 'rewards/meter/std': 0.3643706738948822, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5299999713897705, 'rewards/judge_quality/std': 0.2620250880718231, 'rewards/repeat_penalty/mean': 0.7450039982795715, 'rewards/repeat_penalty/std': 0.27974092960357666, 'rewards/near_duplicate_penalty/mean': 0.919653058052063, 'rewards/near_duplicate_penalty/std': 0.06324919313192368, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.8795064687728882, 'rewards/distinct_2/std': 0.16224141418933868, 'rewards/total_composite/mean': 0.3312518000602722, 'rewards/total_composite/std': 0.2662372589111328, 'reward': 0.3312518000602722, 'reward_std': 0.2662372589111328, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1578971892595291, 'sampling/sampling_logp_difference/max': 1.4733262062072754, 'sampling/importance_sampling_ratio/min': 0.22916197776794434, 'sampling/importance_sampling_ratio/mean': 1.0007827281951904, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0910021141171455, 'clip_ratio/low_mean': 0.05578818405047059, 'clip_ratio/low_min': 0.05578818405047059, 'clip_ratio/high_mean': 0.07909977715462446, 'clip_ratio/high_max': 0.07909977715462446, 'clip_ratio/region_mean': 0.13488796120509505, 'reward_total_mean': 0.3312518000602722, 'reward_meter_mean': 0.6561762094497681, 'reward_meter_std': 0.3643706738948822, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7450039982795715, 'reward_repeat_penalty_std': 0.27974092960357666, 'reward_near_duplicate_penalty_mean': 0.919653058052063, 'reward_near_duplicate_penalty_std': 0.06324919313192368, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.8795064687728882, 'reward_distinct_2_std': 0.16224141418933868, 'reward_judge_quality_mean': 0.5299999713897705, 'reward_judge_quality_std': 0.2620250880718231, 'reward_total_composite_mean': 0.3312518000602722, 'reward_total_composite_std': 0.2662372589111328, 'epoch': 0.02}

 12%|█▏        | 411/3300 [1:34:17<8:54:12, 11.09s/it]INFO 04-12 12:20:29 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 412/3300 [1:34:27<8:31:31, 10.63s/it]2026-04-12 12:20:39,179 | INFO | train_grpo_train | metrics_logged mode=train step=412

                                                      
{'loss': 0.1063, 'grad_norm': 15.938389778137207, 'learning_rate': 8.754545454545456e-06, 'num_tokens': 812424.0, 'completions/mean_length': 59.625, 'completions/min_length': 54.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.625, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.7961858510971069, 'rewards/meter/std': 0.29951679706573486, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5837500095367432, 'rewards/judge_quality/std': 0.19190306961536407, 'rewards/repeat_penalty/mean': 0.9583498239517212, 'rewards/repeat_penalty/std': 0.05034941807389259, 'rewards/near_duplicate_penalty/mean': 0.9699505567550659, 'rewards/near_duplicate_penalty/std': 0.023068485781550407, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9875930547714233, 'rewards/distinct_2/std': 0.03509214147925377, 'rewards/total_composite/mean': 0.4376469850540161, 'rewards/total_composite/std': 0.17945703864097595, 'reward': 0.4376469850540161, 'reward_std': 0.17945702373981476, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1549452245235443, 'sampling/sampling_logp_difference/max': 2.3777527809143066, 'sampling/importance_sampling_ratio/min': 0.09275878965854645, 'sampling/importance_sampling_ratio/mean': 1.0314337015151978, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8387546688318253, 'clip_ratio/low_mean': 0.06667183618992567, 'clip_ratio/low_min': 0.06667183618992567, 'clip_ratio/high_mean': 0.062450396828353405, 'clip_ratio/high_max': 0.062450396828353405, 'clip_ratio/region_mean': 0.12912223301827908, 'reward_total_mean': 0.4376469850540161, 'reward_meter_mean': 0.7961858510971069, 'reward_meter_std': 0.29951679706573486, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9583498239517212, 'reward_repeat_penalty_std': 0.05034941807389259, 'reward_near_duplicate_penalty_mean': 0.9699505567550659, 'reward_near_duplicate_penalty_std': 0.023068485781550407, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9875930547714233, 'reward_distinct_2_std': 0.03509214147925377, 'reward_judge_quality_mean': 0.5837500095367432, 'reward_judge_quality_std': 0.19190306961536407, 'reward_total_composite_mean': 0.4376469850540161, 'reward_total_composite_std': 0.17945703864097595, 'epoch': 0.02}

 12%|█▏        | 412/3300 [1:34:27<8:31:31, 10.63s/it]INFO 04-12 12:20:39 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 413/3300 [1:34:36<8:10:16, 10.19s/it]2026-04-12 12:20:48,340 | INFO | train_grpo_train | metrics_logged mode=train step=413

                                                      
{'loss': 0.0653, 'grad_norm': 14.789229393005371, 'learning_rate': 8.751515151515151e-06, 'num_tokens': 814096.0, 'completions/mean_length': 51.0, 'completions/min_length': 46.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.0, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.602146565914154, 'rewards/meter/std': 0.3833732306957245, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8650000095367432, 'rewards/judge_quality/std': 0.16801361739635468, 'rewards/repeat_penalty/mean': 0.9620596170425415, 'rewards/repeat_penalty/std': 0.03415054455399513, 'rewards/near_duplicate_penalty/mean': 0.9830785989761353, 'rewards/near_duplicate_penalty/std': 0.013511762022972107, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9785538911819458, 'rewards/distinct_2/std': 0.02968541905283928, 'rewards/total_composite/mean': 0.5035417079925537, 'rewards/total_composite/std': 0.3415467441082001, 'reward': 0.5035417079925537, 'reward_std': 0.3415467143058777, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1885005086660385, 'sampling/sampling_logp_difference/max': 1.6356555223464966, 'sampling/importance_sampling_ratio/min': 0.19482462108135223, 'sampling/importance_sampling_ratio/mean': 1.021082878112793, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2148092314600945, 'clip_ratio/low_mean': 0.07938603404909372, 'clip_ratio/low_min': 0.07938603404909372, 'clip_ratio/high_mean': 0.10323066264390945, 'clip_ratio/high_max': 0.10323066264390945, 'clip_ratio/region_mean': 0.18261669669300318, 'reward_total_mean': 0.5035417079925537, 'reward_meter_mean': 0.602146565914154, 'reward_meter_std': 0.3833732306957245, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9620596170425415, 'reward_repeat_penalty_std': 0.03415054455399513, 'reward_near_duplicate_penalty_mean': 0.9830785989761353, 'reward_near_duplicate_penalty_std': 0.013511762022972107, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9785538911819458, 'reward_distinct_2_std': 0.02968541905283928, 'reward_judge_quality_mean': 0.8650000095367432, 'reward_judge_quality_std': 0.16801361739635468, 'reward_total_composite_mean': 0.5035417079925537, 'reward_total_composite_std': 0.3415467441082001, 'epoch': 0.02}

 13%|█▎        | 413/3300 [1:34:36<8:10:16, 10.19s/it]INFO 04-12 12:20:48 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 414/3300 [1:34:44<7:48:15,  9.74s/it]2026-04-12 12:20:57,013 | INFO | train_grpo_train | metrics_logged mode=train step=414

                                                      
{'loss': 0.0331, 'grad_norm': 19.247013092041016, 'learning_rate': 8.748484848484849e-06, 'num_tokens': 815686.0, 'completions/mean_length': 42.75, 'completions/min_length': 39.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.75, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.9652136564254761, 'rewards/meter/std': 0.05562176927924156, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5949999690055847, 'rewards/judge_quality/std': 0.2424871176481247, 'rewards/repeat_penalty/mean': 0.9602253437042236, 'rewards/repeat_penalty/std': 0.03530919924378395, 'rewards/near_duplicate_penalty/mean': 0.9695789813995361, 'rewards/near_duplicate_penalty/std': 0.02480248548090458, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.990384578704834, 'rewards/distinct_2/std': 0.027196412906050682, 'rewards/total_composite/mean': 0.5271258354187012, 'rewards/total_composite/std': 0.29934781789779663, 'reward': 0.5271258354187012, 'reward_std': 0.29934781789779663, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17100238800048828, 'sampling/sampling_logp_difference/max': 1.5007429122924805, 'sampling/importance_sampling_ratio/min': 0.22296446561813354, 'sampling/importance_sampling_ratio/mean': 1.0069575309753418, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3524346947669983, 'clip_ratio/low_mean': 0.06838395539671183, 'clip_ratio/low_min': 0.06838395539671183, 'clip_ratio/high_mean': 0.08870341815054417, 'clip_ratio/high_max': 0.08870341815054417, 'clip_ratio/region_mean': 0.157087373547256, 'reward_total_mean': 0.5271258354187012, 'reward_meter_mean': 0.9652136564254761, 'reward_meter_std': 0.05562176927924156, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9602253437042236, 'reward_repeat_penalty_std': 0.03530919924378395, 'reward_near_duplicate_penalty_mean': 0.9695789813995361, 'reward_near_duplicate_penalty_std': 0.02480248548090458, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.990384578704834, 'reward_distinct_2_std': 0.027196412906050682, 'reward_judge_quality_mean': 0.5949999690055847, 'reward_judge_quality_std': 0.2424871176481247, 'reward_total_composite_mean': 0.5271258354187012, 'reward_total_composite_std': 0.29934781789779663, 'epoch': 0.02}

 13%|█▎        | 414/3300 [1:34:44<7:48:15,  9.74s/it]INFO 04-12 12:20:57 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 415/3300 [1:34:53<7:35:51,  9.48s/it]2026-04-12 12:21:05,900 | INFO | train_grpo_train | metrics_logged mode=train step=415

                                                      
{'loss': 0.0468, 'grad_norm': 24.992507934570312, 'learning_rate': 8.745454545454546e-06, 'num_tokens': 817352.0, 'completions/mean_length': 42.25, 'completions/min_length': 38.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.25, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.8995040655136108, 'rewards/meter/std': 0.18367518484592438, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.05345224589109421, 'rewards/repeat_penalty/mean': 0.8719308972358704, 'rewards/repeat_penalty/std': 0.17576731741428375, 'rewards/near_duplicate_penalty/mean': 0.9603519439697266, 'rewards/near_duplicate_penalty/std': 0.0411485992372036, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9578272104263306, 'rewards/distinct_2/std': 0.06420254707336426, 'rewards/total_composite/mean': 0.35629498958587646, 'rewards/total_composite/std': 0.0774129331111908, 'reward': 0.35629498958587646, 'reward_std': 0.0774129331111908, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1832866072654724, 'sampling/sampling_logp_difference/max': 2.9659318923950195, 'sampling/importance_sampling_ratio/min': 0.05151244252920151, 'sampling/importance_sampling_ratio/mean': 1.0329174995422363, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2682984620332718, 'clip_ratio/low_mean': 0.1089435750618577, 'clip_ratio/low_min': 0.1089435750618577, 'clip_ratio/high_mean': 0.06746562011539936, 'clip_ratio/high_max': 0.06746562011539936, 'clip_ratio/region_mean': 0.17640919517725706, 'reward_total_mean': 0.35629498958587646, 'reward_meter_mean': 0.8995040655136108, 'reward_meter_std': 0.18367518484592438, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8719308972358704, 'reward_repeat_penalty_std': 0.17576731741428375, 'reward_near_duplicate_penalty_mean': 0.9603519439697266, 'reward_near_duplicate_penalty_std': 0.0411485992372036, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9578272104263306, 'reward_distinct_2_std': 0.06420254707336426, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.05345224589109421, 'reward_total_composite_mean': 0.35629498958587646, 'reward_total_composite_std': 0.0774129331111908, 'epoch': 0.02}

 13%|█▎        | 415/3300 [1:34:53<7:35:51,  9.48s/it]INFO 04-12 12:21:06 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 416/3300 [1:35:02<7:19:20,  9.14s/it]2026-04-12 12:21:14,247 | INFO | train_grpo_train | metrics_logged mode=train step=416

                                                      
{'loss': 0.0143, 'grad_norm': 21.97481918334961, 'learning_rate': 8.742424242424243e-06, 'num_tokens': 818755.0, 'completions/mean_length': 23.375, 'completions/min_length': 20.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.375, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.778793454170227, 'rewards/meter/std': 0.3742762506008148, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9730392098426819, 'rewards/lexical_plausibility/std': 0.07625661045312881, 'rewards/judge_quality/mean': 0.65625, 'rewards/judge_quality/std': 0.3692052662372589, 'rewards/repeat_penalty/mean': 0.7336001992225647, 'rewards/repeat_penalty/std': 0.04393468424677849, 'rewards/near_duplicate_penalty/mean': 0.9781335592269897, 'rewards/near_duplicate_penalty/std': 0.05857957899570465, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4568037986755371, 'rewards/total_composite/std': 0.37264272570610046, 'reward': 0.4568037986755371, 'reward_std': 0.37264275550842285, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18755678832530975, 'sampling/sampling_logp_difference/max': 1.2476906776428223, 'sampling/importance_sampling_ratio/min': 0.28716719150543213, 'sampling/importance_sampling_ratio/mean': 1.0562388896942139, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.339102678000927, 'clip_ratio/low_mean': 0.11233766470104456, 'clip_ratio/low_min': 0.11233766470104456, 'clip_ratio/high_mean': 0.06971153896301985, 'clip_ratio/high_max': 0.06971153896301985, 'clip_ratio/region_mean': 0.1820492036640644, 'reward_total_mean': 0.4568037986755371, 'reward_meter_mean': 0.778793454170227, 'reward_meter_std': 0.3742762506008148, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9730392098426819, 'reward_lexical_plausibility_std': 0.07625661045312881, 'reward_repeat_penalty_mean': 0.7336001992225647, 'reward_repeat_penalty_std': 0.04393468424677849, 'reward_near_duplicate_penalty_mean': 0.9781335592269897, 'reward_near_duplicate_penalty_std': 0.05857957899570465, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.65625, 'reward_judge_quality_std': 0.3692052662372589, 'reward_total_composite_mean': 0.4568037986755371, 'reward_total_composite_std': 0.37264272570610046, 'epoch': 0.02}

 13%|█▎        | 416/3300 [1:35:02<7:19:20,  9.14s/it]INFO 04-12 12:21:14 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 417/3300 [1:35:10<7:03:27,  8.81s/it]2026-04-12 12:21:22,297 | INFO | train_grpo_train | metrics_logged mode=train step=417

                                                      
{'loss': -0.0742, 'grad_norm': 28.410341262817383, 'learning_rate': 8.73939393939394e-06, 'num_tokens': 820087.0, 'completions/mean_length': 24.5, 'completions/min_length': 18.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.5, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.9048581719398499, 'rewards/meter/std': 0.21377499401569366, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8487499952316284, 'rewards/judge_quality/std': 0.20152544975280762, 'rewards/repeat_penalty/mean': 0.7361706495285034, 'rewards/repeat_penalty/std': 0.037580110132694244, 'rewards/near_duplicate_penalty/mean': 0.9815608859062195, 'rewards/near_duplicate_penalty/std': 0.050106823444366455, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.8056210279464722, 'rewards/total_composite/std': 0.27250534296035767, 'reward': 0.8056210279464722, 'reward_std': 0.27250534296035767, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16542938351631165, 'sampling/sampling_logp_difference/max': 1.246114730834961, 'sampling/importance_sampling_ratio/min': 0.2876201272010803, 'sampling/importance_sampling_ratio/mean': 1.018161416053772, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3629020899534225, 'clip_ratio/low_mean': 0.013888888992369175, 'clip_ratio/low_min': 0.013888888992369175, 'clip_ratio/high_mean': 0.10890915151685476, 'clip_ratio/high_max': 0.10890915151685476, 'clip_ratio/region_mean': 0.12279804050922394, 'reward_total_mean': 0.8056210279464722, 'reward_meter_mean': 0.9048581719398499, 'reward_meter_std': 0.21377499401569366, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7361706495285034, 'reward_repeat_penalty_std': 0.037580110132694244, 'reward_near_duplicate_penalty_mean': 0.9815608859062195, 'reward_near_duplicate_penalty_std': 0.050106823444366455, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8487499952316284, 'reward_judge_quality_std': 0.20152544975280762, 'reward_total_composite_mean': 0.8056210279464722, 'reward_total_composite_std': 0.27250534296035767, 'epoch': 0.02}

 13%|█▎        | 417/3300 [1:35:10<7:03:27,  8.81s/it]INFO 04-12 12:21:22 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 418/3300 [1:35:23<8:05:32, 10.11s/it]2026-04-12 12:21:35,430 | INFO | train_grpo_train | metrics_logged mode=train step=418

                                                      
{'loss': 0.0564, 'grad_norm': 5.44776725769043, 'learning_rate': 8.736363636363638e-06, 'num_tokens': 823579.0, 'completions/mean_length': 226.5, 'completions/min_length': 178.0, 'completions/max_length': 290.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 226.5, 'completions/min_terminated_length': 178.0, 'completions/max_terminated_length': 290.0, 'rewards/meter/mean': 0.8244215250015259, 'rewards/meter/std': 0.18453948199748993, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.06408698856830597, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/repeat_penalty/mean': 0.3136889338493347, 'rewards/repeat_penalty/std': 0.24498972296714783, 'rewards/near_duplicate_penalty/mean': 0.8750845193862915, 'rewards/near_duplicate_penalty/std': 0.07377022504806519, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.38020312786102295, 'rewards/distinct_2/std': 0.26061731576919556, 'rewards/total_composite/mean': 0.21593712270259857, 'rewards/total_composite/std': 0.07038213312625885, 'reward': 0.21593712270259857, 'reward_std': 0.07038213312625885, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10993286967277527, 'sampling/sampling_logp_difference/max': 3.2446558475494385, 'sampling/importance_sampling_ratio/min': 0.038981977850198746, 'sampling/importance_sampling_ratio/mean': 1.012296199798584, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7208620756864548, 'clip_ratio/low_mean': 0.048527393490076065, 'clip_ratio/low_min': 0.048527393490076065, 'clip_ratio/high_mean': 0.05202482547610998, 'clip_ratio/high_max': 0.05202482547610998, 'clip_ratio/region_mean': 0.10055221896618605, 'reward_total_mean': 0.21593712270259857, 'reward_meter_mean': 0.8244215250015259, 'reward_meter_std': 0.18453948199748993, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.06408698856830597, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.3136889338493347, 'reward_repeat_penalty_std': 0.24498972296714783, 'reward_near_duplicate_penalty_mean': 0.8750845193862915, 'reward_near_duplicate_penalty_std': 0.07377022504806519, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.38020312786102295, 'reward_distinct_2_std': 0.26061731576919556, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.21593712270259857, 'reward_total_composite_std': 0.07038213312625885, 'epoch': 0.02}

 13%|█▎        | 418/3300 [1:35:23<8:05:32, 10.11s/it]INFO 04-12 12:21:35 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 419/3300 [1:35:33<8:06:07, 10.12s/it]2026-04-12 12:21:45,590 | INFO | train_grpo_train | metrics_logged mode=train step=419

                                                      
{'loss': 0.1044, 'grad_norm': 14.743124961853027, 'learning_rate': 8.733333333333333e-06, 'num_tokens': 825519.0, 'completions/mean_length': 73.5, 'completions/min_length': 66.0, 'completions/max_length': 94.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.5, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.7742084860801697, 'rewards/meter/std': 0.32632309198379517, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.42499998211860657, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.9682209491729736, 'rewards/repeat_penalty/std': 0.05635341256856918, 'rewards/near_duplicate_penalty/mean': 0.9899640083312988, 'rewards/near_duplicate_penalty/std': 0.015279407612979412, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9775148034095764, 'rewards/distinct_2/std': 0.04431609809398651, 'rewards/total_composite/mean': 0.32289206981658936, 'rewards/total_composite/std': 0.15256617963314056, 'reward': 0.32289206981658936, 'reward_std': 0.15256617963314056, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1922486424446106, 'sampling/sampling_logp_difference/max': 3.628664493560791, 'sampling/importance_sampling_ratio/min': 0.0265516210347414, 'sampling/importance_sampling_ratio/mean': 1.0239447355270386, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7284061163663864, 'clip_ratio/low_mean': 0.07383543532341719, 'clip_ratio/low_min': 0.07383543532341719, 'clip_ratio/high_mean': 0.07525335624814034, 'clip_ratio/high_max': 0.07525335624814034, 'clip_ratio/region_mean': 0.14908879157155752, 'reward_total_mean': 0.32289206981658936, 'reward_meter_mean': 0.7742084860801697, 'reward_meter_std': 0.32632309198379517, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9682209491729736, 'reward_repeat_penalty_std': 0.05635341256856918, 'reward_near_duplicate_penalty_mean': 0.9899640083312988, 'reward_near_duplicate_penalty_std': 0.015279407612979412, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9775148034095764, 'reward_distinct_2_std': 0.04431609809398651, 'reward_judge_quality_mean': 0.42499998211860657, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.32289206981658936, 'reward_total_composite_std': 0.15256617963314056, 'epoch': 0.02}

 13%|█▎        | 419/3300 [1:35:33<8:06:07, 10.12s/it]INFO 04-12 12:21:45 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 420/3300 [1:35:42<7:51:10,  9.82s/it]2026-04-12 12:21:54,688 | INFO | train_grpo_train | metrics_logged mode=train step=420

                                                      
{'loss': 0.0388, 'grad_norm': 15.905547142028809, 'learning_rate': 8.73030303030303e-06, 'num_tokens': 827197.0, 'completions/mean_length': 29.75, 'completions/min_length': 28.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.75, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.5912508964538574, 'rewards/meter/std': 0.28881779313087463, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.6262500286102295, 'rewards/judge_quality/std': 0.3325631320476532, 'rewards/repeat_penalty/mean': 0.8878109455108643, 'rewards/repeat_penalty/std': 0.10816412419080734, 'rewards/near_duplicate_penalty/mean': 0.9284840822219849, 'rewards/near_duplicate_penalty/std': 0.07428260147571564, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9542124271392822, 'rewards/distinct_2/std': 0.063445083796978, 'rewards/total_composite/mean': 0.32871633768081665, 'rewards/total_composite/std': 0.25182682275772095, 'reward': 0.32871633768081665, 'reward_std': 0.25182679295539856, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1979738175868988, 'sampling/sampling_logp_difference/max': 2.9579548835754395, 'sampling/importance_sampling_ratio/min': 0.05192500352859497, 'sampling/importance_sampling_ratio/mean': 1.0083683729171753, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7957288324832916, 'clip_ratio/low_mean': 0.118675597012043, 'clip_ratio/low_min': 0.118675597012043, 'clip_ratio/high_mean': 0.020312500186264515, 'clip_ratio/high_max': 0.020312500186264515, 'clip_ratio/region_mean': 0.13898809719830751, 'reward_total_mean': 0.32871633768081665, 'reward_meter_mean': 0.5912508964538574, 'reward_meter_std': 0.28881779313087463, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.8878109455108643, 'reward_repeat_penalty_std': 0.10816412419080734, 'reward_near_duplicate_penalty_mean': 0.9284840822219849, 'reward_near_duplicate_penalty_std': 0.07428260147571564, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9542124271392822, 'reward_distinct_2_std': 0.063445083796978, 'reward_judge_quality_mean': 0.6262500286102295, 'reward_judge_quality_std': 0.3325631320476532, 'reward_total_composite_mean': 0.32871633768081665, 'reward_total_composite_std': 0.25182682275772095, 'epoch': 0.02}

 13%|█▎        | 420/3300 [1:35:42<7:51:10,  9.82s/it]INFO 04-12 12:21:54 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 421/3300 [1:35:52<7:57:35,  9.95s/it]2026-04-12 12:22:04,958 | INFO | train_grpo_train | metrics_logged mode=train step=421

                                                      
{'loss': 0.0307, 'grad_norm': 13.426562309265137, 'learning_rate': 8.727272727272728e-06, 'num_tokens': 828886.0, 'completions/mean_length': 49.125, 'completions/min_length': 46.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.125, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.9775240421295166, 'rewards/meter/std': 0.01793941855430603, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4412500262260437, 'rewards/judge_quality/std': 0.1829470992088318, 'rewards/repeat_penalty/mean': 0.9458363056182861, 'rewards/repeat_penalty/std': 0.0516032837331295, 'rewards/near_duplicate_penalty/mean': 0.9621745347976685, 'rewards/near_duplicate_penalty/std': 0.026803744956851006, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9825174808502197, 'rewards/distinct_2/std': 0.03237133473157883, 'rewards/total_composite/mean': 0.4323609471321106, 'rewards/total_composite/std': 0.1838994324207306, 'reward': 0.4323609471321106, 'reward_std': 0.1838994324207306, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14373093843460083, 'sampling/sampling_logp_difference/max': 1.046630859375, 'sampling/importance_sampling_ratio/min': 0.35111871361732483, 'sampling/importance_sampling_ratio/mean': 1.027567744255066, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.535805955529213, 'clip_ratio/low_mean': 0.07628573989495635, 'clip_ratio/low_min': 0.07628573989495635, 'clip_ratio/high_mean': 0.04864583257585764, 'clip_ratio/high_max': 0.04864583257585764, 'clip_ratio/region_mean': 0.12493157247081399, 'reward_total_mean': 0.4323609471321106, 'reward_meter_mean': 0.9775240421295166, 'reward_meter_std': 0.01793941855430603, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9458363056182861, 'reward_repeat_penalty_std': 0.0516032837331295, 'reward_near_duplicate_penalty_mean': 0.9621745347976685, 'reward_near_duplicate_penalty_std': 0.026803744956851006, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9825174808502197, 'reward_distinct_2_std': 0.03237133473157883, 'reward_judge_quality_mean': 0.4412500262260437, 'reward_judge_quality_std': 0.1829470992088318, 'reward_total_composite_mean': 0.4323609471321106, 'reward_total_composite_std': 0.1838994324207306, 'epoch': 0.02}

 13%|█▎        | 421/3300 [1:35:52<7:57:35,  9.95s/it]INFO 04-12 12:22:05 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 422/3300 [1:36:01<7:43:33,  9.66s/it]2026-04-12 12:22:13,949 | INFO | train_grpo_train | metrics_logged mode=train step=422

                                                      
{'loss': 0.0263, 'grad_norm': 13.846668243408203, 'learning_rate': 8.724242424242425e-06, 'num_tokens': 830568.0, 'completions/mean_length': 45.25, 'completions/min_length': 42.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.25, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.754814088344574, 'rewards/meter/std': 0.3215979337692261, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8487499952316284, 'rewards/judge_quality/std': 0.20152544975280762, 'rewards/repeat_penalty/mean': 0.9346069693565369, 'rewards/repeat_penalty/std': 0.18275511264801025, 'rewards/near_duplicate_penalty/mean': 0.9788210988044739, 'rewards/near_duplicate_penalty/std': 0.057720430195331573, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9711538553237915, 'rewards/distinct_2/std': 0.0815892368555069, 'rewards/total_composite/mean': 0.6270456910133362, 'rewards/total_composite/std': 0.3111444413661957, 'reward': 0.6270456910133362, 'reward_std': 0.3111444115638733, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13949206471443176, 'sampling/sampling_logp_difference/max': 2.811610221862793, 'sampling/importance_sampling_ratio/min': 0.06010812893509865, 'sampling/importance_sampling_ratio/mean': 1.0061589479446411, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7933690845966339, 'clip_ratio/low_mean': 0.04892857186496258, 'clip_ratio/low_min': 0.04892857186496258, 'clip_ratio/high_mean': 0.07165032811462879, 'clip_ratio/high_max': 0.07165032811462879, 'clip_ratio/region_mean': 0.12057889997959137, 'reward_total_mean': 0.6270456910133362, 'reward_meter_mean': 0.754814088344574, 'reward_meter_std': 0.3215979337692261, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9346069693565369, 'reward_repeat_penalty_std': 0.18275511264801025, 'reward_near_duplicate_penalty_mean': 0.9788210988044739, 'reward_near_duplicate_penalty_std': 0.057720430195331573, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9711538553237915, 'reward_distinct_2_std': 0.0815892368555069, 'reward_judge_quality_mean': 0.8487499952316284, 'reward_judge_quality_std': 0.20152544975280762, 'reward_total_composite_mean': 0.6270456910133362, 'reward_total_composite_std': 0.3111444413661957, 'epoch': 0.02}

 13%|█▎        | 422/3300 [1:36:01<7:43:33,  9.66s/it]INFO 04-12 12:22:14 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 423/3300 [1:36:10<7:32:01,  9.43s/it]2026-04-12 12:22:22,828 | INFO | train_grpo_train | metrics_logged mode=train step=423

                                                      
{'loss': 0.041, 'grad_norm': 10.513765335083008, 'learning_rate': 8.72121212121212e-06, 'num_tokens': 832063.0, 'completions/mean_length': 41.875, 'completions/min_length': 38.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.875, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.437856525182724, 'rewards/meter/std': 0.42093050479888916, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5387499928474426, 'rewards/judge_quality/std': 0.3358757197856903, 'rewards/repeat_penalty/mean': 0.971872091293335, 'rewards/repeat_penalty/std': 0.06182078272104263, 'rewards/near_duplicate_penalty/mean': 0.9835901856422424, 'rewards/near_duplicate_penalty/std': 0.029320599511265755, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9871795177459717, 'rewards/distinct_2/std': 0.03626188635826111, 'rewards/total_composite/mean': 0.20321273803710938, 'rewards/total_composite/std': 0.27882811427116394, 'reward': 0.20321273803710938, 'reward_std': 0.27882811427116394, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1514754593372345, 'sampling/sampling_logp_difference/max': 2.110161781311035, 'sampling/importance_sampling_ratio/min': 0.12121835350990295, 'sampling/importance_sampling_ratio/mean': 1.0200660228729248, 'sampling/importance_sampling_ratio/max': 1.8916045427322388, 'entropy': 1.2396974563598633, 'clip_ratio/low_mean': 0.09915817994624376, 'clip_ratio/low_min': 0.09915817994624376, 'clip_ratio/high_mean': 0.02738095261156559, 'clip_ratio/high_max': 0.02738095261156559, 'clip_ratio/region_mean': 0.12653913255780935, 'reward_total_mean': 0.20321273803710938, 'reward_meter_mean': 0.437856525182724, 'reward_meter_std': 0.42093050479888916, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.971872091293335, 'reward_repeat_penalty_std': 0.06182078272104263, 'reward_near_duplicate_penalty_mean': 0.9835901856422424, 'reward_near_duplicate_penalty_std': 0.029320599511265755, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9871795177459717, 'reward_distinct_2_std': 0.03626188635826111, 'reward_judge_quality_mean': 0.5387499928474426, 'reward_judge_quality_std': 0.3358757197856903, 'reward_total_composite_mean': 0.20321273803710938, 'reward_total_composite_std': 0.27882811427116394, 'epoch': 0.02}

 13%|█▎        | 423/3300 [1:36:10<7:32:01,  9.43s/it]INFO 04-12 12:22:23 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 424/3300 [1:36:20<7:36:05,  9.52s/it]2026-04-12 12:22:32,547 | INFO | train_grpo_train | metrics_logged mode=train step=424

                                                      
{'loss': 0.0594, 'grad_norm': 14.94465160369873, 'learning_rate': 8.71818181818182e-06, 'num_tokens': 833716.0, 'completions/mean_length': 56.625, 'completions/min_length': 48.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.625, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.8263965249061584, 'rewards/meter/std': 0.17918136715888977, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.05345224589109421, 'rewards/repeat_penalty/mean': 0.9450795650482178, 'rewards/repeat_penalty/std': 0.07274159044027328, 'rewards/near_duplicate_penalty/mean': 0.970674991607666, 'rewards/near_duplicate_penalty/std': 0.027759015560150146, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9725275039672852, 'rewards/distinct_2/std': 0.050869233906269073, 'rewards/total_composite/mean': 0.3288799226284027, 'rewards/total_composite/std': 0.08168979734182358, 'reward': 0.3288799226284027, 'reward_std': 0.08168978989124298, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20187003910541534, 'sampling/sampling_logp_difference/max': 1.8711872100830078, 'sampling/importance_sampling_ratio/min': 0.15394079685211182, 'sampling/importance_sampling_ratio/mean': 1.0487182140350342, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1094235479831696, 'clip_ratio/low_mean': 0.0804588831961155, 'clip_ratio/low_min': 0.0804588831961155, 'clip_ratio/high_mean': 0.08976180106401443, 'clip_ratio/high_max': 0.08976180106401443, 'clip_ratio/region_mean': 0.17022068426012993, 'reward_total_mean': 0.3288799226284027, 'reward_meter_mean': 0.8263965249061584, 'reward_meter_std': 0.17918136715888977, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9450795650482178, 'reward_repeat_penalty_std': 0.07274159044027328, 'reward_near_duplicate_penalty_mean': 0.970674991607666, 'reward_near_duplicate_penalty_std': 0.027759015560150146, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9725275039672852, 'reward_distinct_2_std': 0.050869233906269073, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.05345224589109421, 'reward_total_composite_mean': 0.3288799226284027, 'reward_total_composite_std': 0.08168979734182358, 'epoch': 0.02}

 13%|█▎        | 424/3300 [1:36:20<7:36:05,  9.52s/it]INFO 04-12 12:22:32 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 425/3300 [1:36:28<7:20:46,  9.20s/it]2026-04-12 12:22:41,005 | INFO | train_grpo_train | metrics_logged mode=train step=425

                                                      
{'loss': 0.1842, 'grad_norm': 21.431835174560547, 'learning_rate': 8.715151515151515e-06, 'num_tokens': 835093.0, 'completions/mean_length': 26.125, 'completions/min_length': 21.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.125, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.8926048278808594, 'rewards/meter/std': 0.26264405250549316, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.44999998807907104, 'rewards/judge_quality/std': 0.22677867114543915, 'rewards/repeat_penalty/mean': 0.78125, 'rewards/repeat_penalty/std': 0.0883883461356163, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3880215287208557, 'rewards/total_composite/std': 0.273574560880661, 'reward': 0.3880215287208557, 'reward_std': 0.273574560880661, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1933649182319641, 'sampling/sampling_logp_difference/max': 1.5909218788146973, 'sampling/importance_sampling_ratio/min': 0.20373770594596863, 'sampling/importance_sampling_ratio/mean': 0.9922532439231873, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4773532301187515, 'clip_ratio/low_mean': 0.06284489994868636, 'clip_ratio/low_min': 0.06284489994868636, 'clip_ratio/high_mean': 0.10304631106555462, 'clip_ratio/high_max': 0.10304631106555462, 'clip_ratio/region_mean': 0.16589121101424098, 'reward_total_mean': 0.3880215287208557, 'reward_meter_mean': 0.8926048278808594, 'reward_meter_std': 0.26264405250549316, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.78125, 'reward_repeat_penalty_std': 0.0883883461356163, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.44999998807907104, 'reward_judge_quality_std': 0.22677867114543915, 'reward_total_composite_mean': 0.3880215287208557, 'reward_total_composite_std': 0.273574560880661, 'epoch': 0.02}

 13%|█▎        | 425/3300 [1:36:28<7:20:46,  9.20s/it]INFO 04-12 12:22:41 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 426/3300 [1:36:38<7:19:19,  9.17s/it]2026-04-12 12:22:50,118 | INFO | train_grpo_train | metrics_logged mode=train step=426

                                                      
{'loss': 0.0542, 'grad_norm': 14.11231517791748, 'learning_rate': 8.712121212121212e-06, 'num_tokens': 836620.0, 'completions/mean_length': 41.875, 'completions/min_length': 30.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.875, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.6287662386894226, 'rewards/meter/std': 0.38475966453552246, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5625, 'rewards/judge_quality/std': 0.2282073199748993, 'rewards/repeat_penalty/mean': 0.9871855974197388, 'rewards/repeat_penalty/std': 0.02491079457104206, 'rewards/near_duplicate_penalty/mean': 0.9963431358337402, 'rewards/near_duplicate_penalty/std': 0.005119704641401768, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9908424615859985, 'rewards/distinct_2/std': 0.025901345536112785, 'rewards/total_composite/mean': 0.3885602355003357, 'rewards/total_composite/std': 0.33843523263931274, 'reward': 0.3885602355003357, 'reward_std': 0.33843523263931274, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17799237370491028, 'sampling/sampling_logp_difference/max': 1.22430419921875, 'sampling/importance_sampling_ratio/min': 0.29396215081214905, 'sampling/importance_sampling_ratio/mean': 1.0474486351013184, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9032953679561615, 'clip_ratio/low_mean': 0.10399044491350651, 'clip_ratio/low_min': 0.10399044491350651, 'clip_ratio/high_mean': 0.09663469530642033, 'clip_ratio/high_max': 0.09663469530642033, 'clip_ratio/region_mean': 0.20062514021992683, 'reward_total_mean': 0.3885602355003357, 'reward_meter_mean': 0.6287662386894226, 'reward_meter_std': 0.38475966453552246, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9871855974197388, 'reward_repeat_penalty_std': 0.02491079457104206, 'reward_near_duplicate_penalty_mean': 0.9963431358337402, 'reward_near_duplicate_penalty_std': 0.005119704641401768, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9908424615859985, 'reward_distinct_2_std': 0.025901345536112785, 'reward_judge_quality_mean': 0.5625, 'reward_judge_quality_std': 0.2282073199748993, 'reward_total_composite_mean': 0.3885602355003357, 'reward_total_composite_std': 0.33843523263931274, 'epoch': 0.02}

 13%|█▎        | 426/3300 [1:36:38<7:19:19,  9.17s/it]INFO 04-12 12:22:50 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 427/3300 [1:36:52<8:34:44, 10.75s/it]2026-04-12 12:23:04,546 | INFO | train_grpo_train | metrics_logged mode=train step=427

                                                      
{'loss': -0.1547, 'grad_norm': 3.417048215866089, 'learning_rate': 8.70909090909091e-06, 'num_tokens': 838759.0, 'completions/mean_length': 212.375, 'completions/min_length': 102.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 112.5, 'completions/min_terminated_length': 102.0, 'completions/max_terminated_length': 125.0, 'rewards/meter/mean': 0.6894584894180298, 'rewards/meter/std': 0.3203989267349243, 'rewards/count_adherence/mean': 0.6875, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9561769366264343, 'rewards/lexical_plausibility/std': 0.12395032495260239, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.16903084516525269, 'rewards/repeat_penalty/mean': 0.8323366045951843, 'rewards/repeat_penalty/std': 0.24232245981693268, 'rewards/near_duplicate_penalty/mean': 0.9581027626991272, 'rewards/near_duplicate_penalty/std': 0.07691756635904312, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9047919511795044, 'rewards/distinct_2/std': 0.18057399988174438, 'rewards/total_composite/mean': 0.14915218949317932, 'rewards/total_composite/std': 0.12537094950675964, 'reward': 0.14915218949317932, 'reward_std': 0.12537094950675964, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1877783089876175, 'sampling/sampling_logp_difference/max': 1.2086000442504883, 'sampling/importance_sampling_ratio/min': 0.298615038394928, 'sampling/importance_sampling_ratio/mean': 1.037097692489624, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.444474697113037, 'clip_ratio/low_mean': 0.046176095493137836, 'clip_ratio/low_min': 0.046176095493137836, 'clip_ratio/high_mean': 0.0699781347066164, 'clip_ratio/high_max': 0.0699781347066164, 'clip_ratio/region_mean': 0.11615423019975424, 'reward_total_mean': 0.14915218949317932, 'reward_meter_mean': 0.6894584894180298, 'reward_meter_std': 0.3203989267349243, 'reward_count_adherence_mean': 0.6875, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9561769366264343, 'reward_lexical_plausibility_std': 0.12395032495260239, 'reward_repeat_penalty_mean': 0.8323366045951843, 'reward_repeat_penalty_std': 0.24232245981693268, 'reward_near_duplicate_penalty_mean': 0.9581027626991272, 'reward_near_duplicate_penalty_std': 0.07691756635904312, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9047919511795044, 'reward_distinct_2_std': 0.18057399988174438, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.16903084516525269, 'reward_total_composite_mean': 0.14915218949317932, 'reward_total_composite_std': 0.12537094950675964, 'epoch': 0.02}

 13%|█▎        | 427/3300 [1:36:52<8:34:44, 10.75s/it]INFO 04-12 12:23:04 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 428/3300 [1:37:02<8:23:28, 10.52s/it]2026-04-12 12:23:14,524 | INFO | train_grpo_train | metrics_logged mode=train step=428

                                                      
{'loss': 0.0198, 'grad_norm': 17.723371505737305, 'learning_rate': 8.706060606060607e-06, 'num_tokens': 840302.0, 'completions/mean_length': 29.875, 'completions/min_length': 27.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.875, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.6467967629432678, 'rewards/meter/std': 0.32505813241004944, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.45499998331069946, 'rewards/judge_quality/std': 0.3050995171070099, 'rewards/repeat_penalty/mean': 0.9818909764289856, 'rewards/repeat_penalty/std': 0.029357997700572014, 'rewards/near_duplicate_penalty/mean': 0.9818909764289856, 'rewards/near_duplicate_penalty/std': 0.029357997700572014, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.29642394185066223, 'rewards/total_composite/std': 0.17340131103992462, 'reward': 0.29642394185066223, 'reward_std': 0.17340131103992462, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1463174819946289, 'sampling/sampling_logp_difference/max': 1.5079574584960938, 'sampling/importance_sampling_ratio/min': 0.2213616520166397, 'sampling/importance_sampling_ratio/mean': 1.028766393661499, 'sampling/importance_sampling_ratio/max': 1.8806463479995728, 'entropy': 1.010727547109127, 'clip_ratio/low_mean': 0.020978009328246117, 'clip_ratio/low_min': 0.020978009328246117, 'clip_ratio/high_mean': 0.10347975790500641, 'clip_ratio/high_max': 0.10347975790500641, 'clip_ratio/region_mean': 0.12445776723325253, 'reward_total_mean': 0.29642394185066223, 'reward_meter_mean': 0.6467967629432678, 'reward_meter_std': 0.32505813241004944, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9818909764289856, 'reward_repeat_penalty_std': 0.029357997700572014, 'reward_near_duplicate_penalty_mean': 0.9818909764289856, 'reward_near_duplicate_penalty_std': 0.029357997700572014, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.45499998331069946, 'reward_judge_quality_std': 0.3050995171070099, 'reward_total_composite_mean': 0.29642394185066223, 'reward_total_composite_std': 0.17340131103992462, 'epoch': 0.02}

 13%|█▎        | 428/3300 [1:37:02<8:23:28, 10.52s/it]INFO 04-12 12:23:14 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 429/3300 [1:37:11<7:59:27, 10.02s/it]2026-04-12 12:23:23,385 | INFO | train_grpo_train | metrics_logged mode=train step=429

                                                      
{'loss': 0.0388, 'grad_norm': 19.33844566345215, 'learning_rate': 8.703030303030304e-06, 'num_tokens': 841669.0, 'completions/mean_length': 19.875, 'completions/min_length': 18.0, 'completions/max_length': 21.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.875, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 21.0, 'rewards/meter/mean': 0.737424373626709, 'rewards/meter/std': 0.4435030221939087, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.8652982115745544, 'rewards/lexical_plausibility/std': 0.15936560928821564, 'rewards/judge_quality/mean': 0.4300000071525574, 'rewards/judge_quality/std': 0.3481789529323578, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3620902895927429, 'rewards/total_composite/std': 0.39068546891212463, 'reward': 0.3620902895927429, 'reward_std': 0.39068543910980225, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16863738000392914, 'sampling/sampling_logp_difference/max': 1.6787558794021606, 'sampling/importance_sampling_ratio/min': 0.18660598993301392, 'sampling/importance_sampling_ratio/mean': 1.0252758264541626, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8704716861248016, 'clip_ratio/low_mean': 0.13951545860618353, 'clip_ratio/low_min': 0.13951545860618353, 'clip_ratio/high_mean': 0.054464287124574184, 'clip_ratio/high_max': 0.054464287124574184, 'clip_ratio/region_mean': 0.1939797457307577, 'reward_total_mean': 0.3620902895927429, 'reward_meter_mean': 0.737424373626709, 'reward_meter_std': 0.4435030221939087, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.8652982115745544, 'reward_lexical_plausibility_std': 0.15936560928821564, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4300000071525574, 'reward_judge_quality_std': 0.3481789529323578, 'reward_total_composite_mean': 0.3620902895927429, 'reward_total_composite_std': 0.39068546891212463, 'epoch': 0.02}

 13%|█▎        | 429/3300 [1:37:11<7:59:27, 10.02s/it]INFO 04-12 12:23:23 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 430/3300 [1:37:21<7:57:03,  9.97s/it]2026-04-12 12:23:33,245 | INFO | train_grpo_train | metrics_logged mode=train step=430

                                                      
{'loss': 0.1789, 'grad_norm': 17.62770652770996, 'learning_rate': 8.700000000000001e-06, 'num_tokens': 843436.0, 'completions/mean_length': 56.875, 'completions/min_length': 50.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.875, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.5594629049301147, 'rewards/meter/std': 0.3411722183227539, '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/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.5550000071525574, 'rewards/judge_quality/std': 0.22790977358818054, 'rewards/repeat_penalty/mean': 0.9628264307975769, 'rewards/repeat_penalty/std': 0.03612314537167549, 'rewards/near_duplicate_penalty/mean': 0.9628264307975769, 'rewards/near_duplicate_penalty/std': 0.03612314537167549, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3280724287033081, 'rewards/total_composite/std': 0.22638551890850067, 'reward': 0.3280724287033081, 'reward_std': 0.22638550400733948, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18308889865875244, 'sampling/sampling_logp_difference/max': 2.3320999145507812, 'sampling/importance_sampling_ratio/min': 0.09709165245294571, 'sampling/importance_sampling_ratio/mean': 1.0079594850540161, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0582946464419365, 'clip_ratio/low_mean': 0.08709927462041378, 'clip_ratio/low_min': 0.08709927462041378, 'clip_ratio/high_mean': 0.053845726884901524, 'clip_ratio/high_max': 0.053845726884901524, 'clip_ratio/region_mean': 0.1409450015053153, 'reward_total_mean': 0.3280724287033081, 'reward_meter_mean': 0.5594629049301147, 'reward_meter_std': 0.3411722183227539, '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_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.9628264307975769, 'reward_repeat_penalty_std': 0.03612314537167549, 'reward_near_duplicate_penalty_mean': 0.9628264307975769, 'reward_near_duplicate_penalty_std': 0.03612314537167549, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5550000071525574, 'reward_judge_quality_std': 0.22790977358818054, 'reward_total_composite_mean': 0.3280724287033081, 'reward_total_composite_std': 0.22638551890850067, 'epoch': 0.02}

 13%|█▎        | 430/3300 [1:37:21<7:57:03,  9.97s/it]INFO 04-12 12:23:33 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 431/3300 [1:37:35<8:54:29, 11.18s/it]2026-04-12 12:23:47,234 | INFO | train_grpo_train | metrics_logged mode=train step=431

                                                      
{'loss': -0.1857, 'grad_norm': 3.314324140548706, 'learning_rate': 8.696969696969699e-06, 'num_tokens': 846166.0, 'completions/mean_length': 279.25, 'completions/min_length': 179.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 201.6666717529297, 'completions/min_terminated_length': 179.0, 'completions/max_terminated_length': 231.0, 'rewards/meter/mean': 0.7354225516319275, 'rewards/meter/std': 0.42297622561454773, 'rewards/count_adherence/mean': 0.8611111044883728, 'rewards/count_adherence/std': 0.17568209767341614, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9780093431472778, 'rewards/lexical_plausibility/std': 0.04118134826421738, 'rewards/judge_quality/mean': 0.13750000298023224, 'rewards/judge_quality/std': 0.09910312294960022, 'rewards/repeat_penalty/mean': 0.4406975507736206, 'rewards/repeat_penalty/std': 0.4031829237937927, 'rewards/near_duplicate_penalty/mean': 0.8388845920562744, 'rewards/near_duplicate_penalty/std': 0.16234926879405975, 'rewards/opening_diversity/mean': 0.8854166865348816, 'rewards/opening_diversity/std': 0.205226868391037, 'rewards/distinct_2/mean': 0.5116329193115234, 'rewards/distinct_2/std': 0.39123836159706116, 'rewards/total_composite/mean': 0.10899841785430908, 'rewards/total_composite/std': 0.09921231120824814, 'reward': 0.10899841785430908, 'reward_std': 0.09921231120824814, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10823307186365128, 'sampling/sampling_logp_difference/max': 1.6192960739135742, 'sampling/importance_sampling_ratio/min': 0.1980380415916443, 'sampling/importance_sampling_ratio/mean': 1.0231510400772095, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8005856201052666, 'clip_ratio/low_mean': 0.008207070641219616, 'clip_ratio/low_min': 0.008207070641219616, 'clip_ratio/high_mean': 0.07269664574414492, 'clip_ratio/high_max': 0.07269664574414492, 'clip_ratio/region_mean': 0.08090371638536453, 'reward_total_mean': 0.10899841785430908, 'reward_meter_mean': 0.7354225516319275, 'reward_meter_std': 0.42297622561454773, 'reward_count_adherence_mean': 0.8611111044883728, 'reward_count_adherence_std': 0.17568209767341614, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9780093431472778, 'reward_lexical_plausibility_std': 0.04118134826421738, 'reward_repeat_penalty_mean': 0.4406975507736206, 'reward_repeat_penalty_std': 0.4031829237937927, 'reward_near_duplicate_penalty_mean': 0.8388845920562744, 'reward_near_duplicate_penalty_std': 0.16234926879405975, 'reward_opening_diversity_mean': 0.8854166865348816, 'reward_opening_diversity_std': 0.205226868391037, 'reward_distinct_2_mean': 0.5116329193115234, 'reward_distinct_2_std': 0.39123836159706116, 'reward_judge_quality_mean': 0.13750000298023224, 'reward_judge_quality_std': 0.09910312294960022, 'reward_total_composite_mean': 0.10899841785430908, 'reward_total_composite_std': 0.09921231120824814, 'epoch': 0.02}

 13%|█▎        | 431/3300 [1:37:35<8:54:29, 11.18s/it]INFO 04-12 12:23:47 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 432/3300 [1:37:48<9:32:17, 11.97s/it]2026-04-12 12:24:01,068 | INFO | train_grpo_train | metrics_logged mode=train step=432

                                                      
{'loss': -0.0278, 'grad_norm': 1.1596909761428833, 'learning_rate': 8.693939393939394e-06, 'num_tokens': 847706.0, 'completions/mean_length': 329.5, 'completions/min_length': 22.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.625, 'completions/mean_terminated_length': 25.33333396911621, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.3426334261894226, 'rewards/meter/std': 0.46065592765808105, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/lexical_plausibility/mean': 0.810083270072937, 'rewards/lexical_plausibility/std': 0.218026265501976, 'rewards/judge_quality/mean': 0.125, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.08040358871221542, 'rewards/total_composite/std': 0.1297454982995987, 'reward': 0.08040358871221542, 'reward_std': 0.1297454982995987, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.26203233003616333, 'sampling/sampling_logp_difference/max': 1.2051019668579102, 'sampling/importance_sampling_ratio/min': 0.29966145753860474, 'sampling/importance_sampling_ratio/mean': 1.066463589668274, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0534523129463196, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.06335227377712727, 'clip_ratio/high_max': 0.06335227377712727, 'clip_ratio/region_mean': 0.06335227377712727, 'reward_total_mean': 0.08040358871221542, 'reward_meter_mean': 0.3426334261894226, 'reward_meter_std': 0.46065592765808105, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_lexical_plausibility_mean': 0.810083270072937, 'reward_lexical_plausibility_std': 0.218026265501976, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.125, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.08040358871221542, 'reward_total_composite_std': 0.1297454982995987, 'epoch': 0.02}

 13%|█▎        | 432/3300 [1:37:49<9:32:17, 11.97s/it]INFO 04-12 12:24:01 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 433/3300 [1:37:59<9:04:43, 11.40s/it]2026-04-12 12:24:11,125 | INFO | train_grpo_train | metrics_logged mode=train step=433

                                                      
{'loss': 0.0604, 'grad_norm': 13.389408111572266, 'learning_rate': 8.690909090909091e-06, 'num_tokens': 849480.0, 'completions/mean_length': 41.75, 'completions/min_length': 36.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.75, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.9307582378387451, 'rewards/meter/std': 0.09821143746376038, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.44624999165534973, 'rewards/judge_quality/std': 0.19668231904506683, 'rewards/repeat_penalty/mean': 0.9533466696739197, 'rewards/repeat_penalty/std': 0.11704007536172867, 'rewards/near_duplicate_penalty/mean': 0.9920774698257446, 'rewards/near_duplicate_penalty/std': 0.012791228480637074, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9886877536773682, 'rewards/distinct_2/std': 0.031995780766010284, 'rewards/total_composite/mean': 0.399087131023407, 'rewards/total_composite/std': 0.10525088757276535, 'reward': 0.399087131023407, 'reward_std': 0.10525088757276535, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14620551466941833, 'sampling/sampling_logp_difference/max': 2.6977453231811523, 'sampling/importance_sampling_ratio/min': 0.06735721230506897, 'sampling/importance_sampling_ratio/mean': 1.0105962753295898, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9248259142041206, 'clip_ratio/low_mean': 0.0850168839097023, 'clip_ratio/low_min': 0.0850168839097023, 'clip_ratio/high_mean': 0.05247699562460184, 'clip_ratio/high_max': 0.05247699562460184, 'clip_ratio/region_mean': 0.13749387953430414, 'reward_total_mean': 0.399087131023407, 'reward_meter_mean': 0.9307582378387451, 'reward_meter_std': 0.09821143746376038, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9533466696739197, 'reward_repeat_penalty_std': 0.11704007536172867, 'reward_near_duplicate_penalty_mean': 0.9920774698257446, 'reward_near_duplicate_penalty_std': 0.012791228480637074, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9886877536773682, 'reward_distinct_2_std': 0.031995780766010284, 'reward_judge_quality_mean': 0.44624999165534973, 'reward_judge_quality_std': 0.19668231904506683, 'reward_total_composite_mean': 0.399087131023407, 'reward_total_composite_std': 0.10525088757276535, 'epoch': 0.02}

 13%|█▎        | 433/3300 [1:37:59<9:04:43, 11.40s/it]INFO 04-12 12:24:11 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 434/3300 [1:38:12<9:33:15, 12.00s/it]2026-04-12 12:24:24,538 | INFO | train_grpo_train | metrics_logged mode=train step=434

                                                      
{'loss': -0.0701, 'grad_norm': 5.030261993408203, 'learning_rate': 8.687878787878789e-06, 'num_tokens': 851100.0, 'completions/mean_length': 105.5, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 47.42857360839844, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.8845380544662476, 'rewards/meter/std': 0.16779212653636932, '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/lexical_plausibility/mean': 0.9858478307723999, 'rewards/lexical_plausibility/std': 0.0400284007191658, 'rewards/judge_quality/mean': 0.48374998569488525, 'rewards/judge_quality/std': 0.25961166620254517, 'rewards/repeat_penalty/mean': 0.9307153224945068, 'rewards/repeat_penalty/std': 0.08318975567817688, 'rewards/near_duplicate_penalty/mean': 0.9793448448181152, 'rewards/near_duplicate_penalty/std': 0.03710789978504181, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9825174808502197, 'rewards/distinct_2/std': 0.03237133473157883, 'rewards/total_composite/mean': 0.40029793977737427, 'rewards/total_composite/std': 0.31994324922561646, 'reward': 0.40029793977737427, 'reward_std': 0.31994324922561646, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16955596208572388, 'sampling/sampling_logp_difference/max': 1.5196118354797363, 'sampling/importance_sampling_ratio/min': 0.21879681944847107, 'sampling/importance_sampling_ratio/mean': 1.0280218124389648, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.387006402015686, 'clip_ratio/low_mean': 0.058102767914533615, 'clip_ratio/low_min': 0.058102767914533615, 'clip_ratio/high_mean': 0.0664657149463892, 'clip_ratio/high_max': 0.0664657149463892, 'clip_ratio/region_mean': 0.12456848286092281, 'reward_total_mean': 0.40029793977737427, 'reward_meter_mean': 0.8845380544662476, 'reward_meter_std': 0.16779212653636932, '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_lexical_plausibility_mean': 0.9858478307723999, 'reward_lexical_plausibility_std': 0.0400284007191658, 'reward_repeat_penalty_mean': 0.9307153224945068, 'reward_repeat_penalty_std': 0.08318975567817688, 'reward_near_duplicate_penalty_mean': 0.9793448448181152, 'reward_near_duplicate_penalty_std': 0.03710789978504181, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9825174808502197, 'reward_distinct_2_std': 0.03237133473157883, 'reward_judge_quality_mean': 0.48374998569488525, 'reward_judge_quality_std': 0.25961166620254517, 'reward_total_composite_mean': 0.40029793977737427, 'reward_total_composite_std': 0.31994324922561646, 'epoch': 0.02}

 13%|█▎        | 434/3300 [1:38:12<9:33:15, 12.00s/it]INFO 04-12 12:24:24 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 435/3300 [1:38:26<10:00:47, 12.58s/it]2026-04-12 12:24:38,467 | INFO | train_grpo_train | metrics_logged mode=train step=435

                                                       
{'loss': -0.0125, 'grad_norm': 5.3134379386901855, 'learning_rate': 8.684848484848486e-06, 'num_tokens': 852750.0, 'completions/mean_length': 105.25, 'completions/min_length': 41.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 47.142860412597656, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.40984004735946655, 'rewards/meter/std': 0.35261011123657227, '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/lexical_plausibility/mean': 0.9995675086975098, 'rewards/lexical_plausibility/std': 0.0012233754387125373, 'rewards/judge_quality/mean': 0.6512500047683716, 'rewards/judge_quality/std': 0.39617595076560974, 'rewards/repeat_penalty/mean': 0.9522122144699097, 'rewards/repeat_penalty/std': 0.11480019986629486, 'rewards/near_duplicate_penalty/mean': 0.9674084186553955, 'rewards/near_duplicate_penalty/std': 0.07194847613573074, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9807692170143127, 'rewards/distinct_2/std': 0.054392825812101364, 'rewards/total_composite/mean': 0.31701648235321045, 'rewards/total_composite/std': 0.36174580454826355, 'reward': 0.31701648235321045, 'reward_std': 0.36174577474594116, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15765489637851715, 'sampling/sampling_logp_difference/max': 1.686666488647461, 'sampling/importance_sampling_ratio/min': 0.1851356476545334, 'sampling/importance_sampling_ratio/mean': 1.0341609716415405, 'sampling/importance_sampling_ratio/max': 1.6873053312301636, 'entropy': 1.1073570996522903, 'clip_ratio/low_mean': 0.07283596880733967, 'clip_ratio/low_min': 0.07283596880733967, 'clip_ratio/high_mean': 0.058390771970152855, 'clip_ratio/high_max': 0.058390771970152855, 'clip_ratio/region_mean': 0.13122674077749252, 'reward_total_mean': 0.31701648235321045, 'reward_meter_mean': 0.40984004735946655, 'reward_meter_std': 0.35261011123657227, '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_lexical_plausibility_mean': 0.9995675086975098, 'reward_lexical_plausibility_std': 0.0012233754387125373, 'reward_repeat_penalty_mean': 0.9522122144699097, 'reward_repeat_penalty_std': 0.11480019986629486, 'reward_near_duplicate_penalty_mean': 0.9674084186553955, 'reward_near_duplicate_penalty_std': 0.07194847613573074, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9807692170143127, 'reward_distinct_2_std': 0.054392825812101364, 'reward_judge_quality_mean': 0.6512500047683716, 'reward_judge_quality_std': 0.39617595076560974, 'reward_total_composite_mean': 0.31701648235321045, 'reward_total_composite_std': 0.36174580454826355, 'epoch': 0.02}

 13%|█▎        | 435/3300 [1:38:26<10:00:47, 12.58s/it]INFO 04-12 12:24:38 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 436/3300 [1:38:41<10:36:36, 13.34s/it]2026-04-12 12:24:53,568 | INFO | train_grpo_train | metrics_logged mode=train step=436

                                                       
{'loss': -0.1395, 'grad_norm': 2.650196075439453, 'learning_rate': 8.681818181818182e-06, 'num_tokens': 854758.0, 'completions/mean_length': 260.0, 'completions/min_length': 99.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 108.80000305175781, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 117.0, 'rewards/meter/mean': 0.7064166069030762, 'rewards/meter/std': 0.3930877447128296, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.13363061845302582, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.9454888105392456, 'rewards/lexical_plausibility/std': 0.08788902312517166, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.16690459847450256, 'rewards/repeat_penalty/mean': 0.8705521821975708, 'rewards/repeat_penalty/std': 0.18051165342330933, 'rewards/near_duplicate_penalty/mean': 0.9721640944480896, 'rewards/near_duplicate_penalty/std': 0.0354461669921875, 'rewards/opening_diversity/mean': 0.979687511920929, 'rewards/opening_diversity/std': 0.03892386704683304, 'rewards/distinct_2/mean': 0.9269729852676392, 'rewards/distinct_2/std': 0.10987544804811478, 'rewards/total_composite/mean': 0.13720133900642395, 'rewards/total_composite/std': 0.15049776434898376, 'reward': 0.13720133900642395, 'reward_std': 0.15049777925014496, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18098576366901398, 'sampling/sampling_logp_difference/max': 1.8638801574707031, 'sampling/importance_sampling_ratio/min': 0.1550697684288025, 'sampling/importance_sampling_ratio/mean': 1.0196309089660645, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0312328487634659, 'clip_ratio/low_mean': 0.03150034788995981, 'clip_ratio/low_min': 0.03150034788995981, 'clip_ratio/high_mean': 0.05556563567370176, 'clip_ratio/high_max': 0.05556563567370176, 'clip_ratio/region_mean': 0.08706598356366158, 'reward_total_mean': 0.13720133900642395, 'reward_meter_mean': 0.7064166069030762, 'reward_meter_std': 0.3930877447128296, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.13363061845302582, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.9454888105392456, 'reward_lexical_plausibility_std': 0.08788902312517166, 'reward_repeat_penalty_mean': 0.8705521821975708, 'reward_repeat_penalty_std': 0.18051165342330933, 'reward_near_duplicate_penalty_mean': 0.9721640944480896, 'reward_near_duplicate_penalty_std': 0.0354461669921875, 'reward_opening_diversity_mean': 0.979687511920929, 'reward_opening_diversity_std': 0.03892386704683304, 'reward_distinct_2_mean': 0.9269729852676392, 'reward_distinct_2_std': 0.10987544804811478, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.16690459847450256, 'reward_total_composite_mean': 0.13720133900642395, 'reward_total_composite_std': 0.15049776434898376, 'epoch': 0.02}

 13%|█▎        | 436/3300 [1:38:41<10:36:36, 13.34s/it]INFO 04-12 12:24:53 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 437/3300 [1:38:56<11:01:11, 13.86s/it]2026-04-12 12:25:08,634 | INFO | train_grpo_train | metrics_logged mode=train step=437

                                                       
{'loss': -0.0699, 'grad_norm': 3.9052555561065674, 'learning_rate': 8.67878787878788e-06, 'num_tokens': 856167.0, 'completions/mean_length': 160.125, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 42.833335876464844, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.5357563495635986, 'rewards/meter/std': 0.4217534363269806, '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/lexical_plausibility/mean': 0.9192672967910767, 'rewards/lexical_plausibility/std': 0.13289739191532135, 'rewards/judge_quality/mean': 0.47999998927116394, 'rewards/judge_quality/std': 0.33903223276138306, 'rewards/repeat_penalty/mean': 0.9827923774719238, 'rewards/repeat_penalty/std': 0.033745668828487396, 'rewards/near_duplicate_penalty/mean': 0.9921927452087402, 'rewards/near_duplicate_penalty/std': 0.01090940646827221, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.990384578704834, 'rewards/distinct_2/std': 0.027196412906050682, 'rewards/total_composite/mean': 0.3074667453765869, 'rewards/total_composite/std': 0.2921159267425537, 'reward': 0.3074667453765869, 'reward_std': 0.2921159267425537, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16886970400810242, 'sampling/sampling_logp_difference/max': 1.2983119487762451, 'sampling/importance_sampling_ratio/min': 0.27299225330352783, 'sampling/importance_sampling_ratio/mean': 0.9962414503097534, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8901057541370392, 'clip_ratio/low_mean': 0.033630953170359135, 'clip_ratio/low_min': 0.033630953170359135, 'clip_ratio/high_mean': 0.0971474926918745, 'clip_ratio/high_max': 0.0971474926918745, 'clip_ratio/region_mean': 0.13077844586223364, 'reward_total_mean': 0.3074667453765869, 'reward_meter_mean': 0.5357563495635986, 'reward_meter_std': 0.4217534363269806, '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_lexical_plausibility_mean': 0.9192672967910767, 'reward_lexical_plausibility_std': 0.13289739191532135, 'reward_repeat_penalty_mean': 0.9827923774719238, 'reward_repeat_penalty_std': 0.033745668828487396, 'reward_near_duplicate_penalty_mean': 0.9921927452087402, 'reward_near_duplicate_penalty_std': 0.01090940646827221, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.990384578704834, 'reward_distinct_2_std': 0.027196412906050682, 'reward_judge_quality_mean': 0.47999998927116394, 'reward_judge_quality_std': 0.33903223276138306, 'reward_total_composite_mean': 0.3074667453765869, 'reward_total_composite_std': 0.2921159267425537, 'epoch': 0.02}

 13%|█▎        | 437/3300 [1:38:56<11:01:11, 13.86s/it]INFO 04-12 12:25:08 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 438/3300 [1:39:12<11:32:29, 14.52s/it]2026-04-12 12:25:24,695 | INFO | train_grpo_train | metrics_logged mode=train step=438

                                                       
{'loss': -0.1876, 'grad_norm': 3.350242853164673, 'learning_rate': 8.675757575757576e-06, 'num_tokens': 859289.0, 'completions/mean_length': 349.25, 'completions/min_length': 232.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 251.60000610351562, 'completions/min_terminated_length': 232.0, 'completions/max_terminated_length': 288.0, 'rewards/meter/mean': 0.5100666284561157, 'rewards/meter/std': 0.3364863097667694, 'rewards/count_adherence/mean': 0.6625000238418579, 'rewards/count_adherence/std': 0.2825268507003784, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/lexical_plausibility/mean': 0.9547035694122314, 'rewards/lexical_plausibility/std': 0.0760425329208374, 'rewards/judge_quality/mean': 0.20000000298023224, 'rewards/judge_quality/std': 0.13093073666095734, 'rewards/repeat_penalty/mean': 0.7465248107910156, 'rewards/repeat_penalty/std': 0.31817758083343506, 'rewards/near_duplicate_penalty/mean': 0.9175324440002441, 'rewards/near_duplicate_penalty/std': 0.10367850214242935, 'rewards/opening_diversity/mean': 0.9545454382896423, 'rewards/opening_diversity/std': 0.0917171910405159, 'rewards/distinct_2/mean': 0.8109216690063477, 'rewards/distinct_2/std': 0.25177526473999023, 'rewards/total_composite/mean': 0.06108936667442322, 'rewards/total_composite/std': 0.08259936422109604, 'reward': 0.06108936667442322, 'reward_std': 0.08259936422109604, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11463236063718796, 'sampling/sampling_logp_difference/max': 2.060795783996582, 'sampling/importance_sampling_ratio/min': 0.12735258042812347, 'sampling/importance_sampling_ratio/mean': 1.0202183723449707, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5843771174550056, 'clip_ratio/low_mean': 0.025493130087852478, 'clip_ratio/low_min': 0.025493130087852478, 'clip_ratio/high_mean': 0.0423896973952651, 'clip_ratio/high_max': 0.0423896973952651, 'clip_ratio/region_mean': 0.06788282748311758, 'reward_total_mean': 0.06108936667442322, 'reward_meter_mean': 0.5100666284561157, 'reward_meter_std': 0.3364863097667694, 'reward_count_adherence_mean': 0.6625000238418579, 'reward_count_adherence_std': 0.2825268507003784, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_lexical_plausibility_mean': 0.9547035694122314, 'reward_lexical_plausibility_std': 0.0760425329208374, 'reward_repeat_penalty_mean': 0.7465248107910156, 'reward_repeat_penalty_std': 0.31817758083343506, 'reward_near_duplicate_penalty_mean': 0.9175324440002441, 'reward_near_duplicate_penalty_std': 0.10367850214242935, 'reward_opening_diversity_mean': 0.9545454382896423, 'reward_opening_diversity_std': 0.0917171910405159, 'reward_distinct_2_mean': 0.8109216690063477, 'reward_distinct_2_std': 0.25177526473999023, 'reward_judge_quality_mean': 0.20000000298023224, 'reward_judge_quality_std': 0.13093073666095734, 'reward_total_composite_mean': 0.06108936667442322, 'reward_total_composite_std': 0.08259936422109604, 'epoch': 0.02}

 13%|█▎        | 438/3300 [1:39:12<11:32:29, 14.52s/it]INFO 04-12 12:25:24 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 439/3300 [1:39:31<12:28:16, 15.69s/it]2026-04-12 12:25:43,130 | INFO | train_grpo_train | metrics_logged mode=train step=439

                                                       
{'loss': -0.099, 'grad_norm': 4.8546142578125, 'learning_rate': 8.672727272727273e-06, 'num_tokens': 860998.0, 'completions/mean_length': 108.625, 'completions/min_length': 46.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 51.000003814697266, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.789451003074646, 'rewards/meter/std': 0.35648372769355774, '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/lexical_plausibility/mean': 0.9670825004577637, 'rewards/lexical_plausibility/std': 0.0931047722697258, 'rewards/judge_quality/mean': 0.4337500035762787, 'rewards/judge_quality/std': 0.23868316411972046, 'rewards/repeat_penalty/mean': 0.8674712181091309, 'rewards/repeat_penalty/std': 0.17557385563850403, 'rewards/near_duplicate_penalty/mean': 0.9874880313873291, 'rewards/near_duplicate_penalty/std': 0.017842702567577362, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9610019326210022, 'rewards/distinct_2/std': 0.05533240735530853, 'rewards/total_composite/mean': 0.3033324182033539, 'rewards/total_composite/std': 0.19120368361473083, 'reward': 0.3033324182033539, 'reward_std': 0.19120366871356964, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15717534720897675, 'sampling/sampling_logp_difference/max': 1.75221848487854, 'sampling/importance_sampling_ratio/min': 0.17338885366916656, 'sampling/importance_sampling_ratio/mean': 1.0023897886276245, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2545001059770584, 'clip_ratio/low_mean': 0.015625, 'clip_ratio/low_min': 0.015625, 'clip_ratio/high_mean': 0.12267677020281553, 'clip_ratio/high_max': 0.12267677020281553, 'clip_ratio/region_mean': 0.13830177020281553, 'reward_total_mean': 0.3033324182033539, 'reward_meter_mean': 0.789451003074646, 'reward_meter_std': 0.35648372769355774, '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_lexical_plausibility_mean': 0.9670825004577637, 'reward_lexical_plausibility_std': 0.0931047722697258, 'reward_repeat_penalty_mean': 0.8674712181091309, 'reward_repeat_penalty_std': 0.17557385563850403, 'reward_near_duplicate_penalty_mean': 0.9874880313873291, 'reward_near_duplicate_penalty_std': 0.017842702567577362, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9610019326210022, 'reward_distinct_2_std': 0.05533240735530853, 'reward_judge_quality_mean': 0.4337500035762787, 'reward_judge_quality_std': 0.23868316411972046, 'reward_total_composite_mean': 0.3033324182033539, 'reward_total_composite_std': 0.19120368361473083, 'epoch': 0.02}

 13%|█▎        | 439/3300 [1:39:31<12:28:16, 15.69s/it]INFO 04-12 12:25:43 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 440/3300 [1:39:44<12:01:18, 15.13s/it]2026-04-12 12:25:56,957 | INFO | train_grpo_train | metrics_logged mode=train step=440

                                                       
{'loss': -0.0349, 'grad_norm': 7.668520927429199, 'learning_rate': 8.66969696969697e-06, 'num_tokens': 862982.0, 'completions/mean_length': 137.0, 'completions/min_length': 73.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 83.42857360839844, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.447530061006546, 'rewards/meter/std': 0.29448914527893066, '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/lexical_plausibility/mean': 0.9737329483032227, 'rewards/lexical_plausibility/std': 0.07429433614015579, 'rewards/judge_quality/mean': 0.5262500047683716, 'rewards/judge_quality/std': 0.35379326343536377, 'rewards/repeat_penalty/mean': 0.9485679864883423, 'rewards/repeat_penalty/std': 0.04207950457930565, 'rewards/near_duplicate_penalty/mean': 0.9766952991485596, 'rewards/near_duplicate_penalty/std': 0.014643006958067417, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9709049463272095, 'rewards/distinct_2/std': 0.03147156536579132, 'rewards/total_composite/mean': 0.26080653071403503, 'rewards/total_composite/std': 0.23689928650856018, 'reward': 0.26080653071403503, 'reward_std': 0.236899271607399, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20429247617721558, 'sampling/sampling_logp_difference/max': 1.795257568359375, 'sampling/importance_sampling_ratio/min': 0.1982724815607071, 'sampling/importance_sampling_ratio/mean': 0.9991369247436523, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8635646998882294, 'clip_ratio/low_mean': 0.06671988312155008, 'clip_ratio/low_min': 0.06671988312155008, 'clip_ratio/high_mean': 0.09525706991553307, 'clip_ratio/high_max': 0.09525706991553307, 'clip_ratio/region_mean': 0.16197695303708315, 'reward_total_mean': 0.26080653071403503, 'reward_meter_mean': 0.447530061006546, 'reward_meter_std': 0.29448914527893066, '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_lexical_plausibility_mean': 0.9737329483032227, 'reward_lexical_plausibility_std': 0.07429433614015579, 'reward_repeat_penalty_mean': 0.9485679864883423, 'reward_repeat_penalty_std': 0.04207950457930565, 'reward_near_duplicate_penalty_mean': 0.9766952991485596, 'reward_near_duplicate_penalty_std': 0.014643006958067417, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9709049463272095, 'reward_distinct_2_std': 0.03147156536579132, 'reward_judge_quality_mean': 0.5262500047683716, 'reward_judge_quality_std': 0.35379326343536377, 'reward_total_composite_mean': 0.26080653071403503, 'reward_total_composite_std': 0.23689928650856018, 'epoch': 0.02}

 13%|█▎        | 440/3300 [1:39:44<12:01:18, 15.13s/it]INFO 04-12 12:25:57 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 441/3300 [1:39:59<11:49:57, 14.90s/it]2026-04-12 12:26:11,313 | INFO | train_grpo_train | metrics_logged mode=train step=441

                                                       
{'loss': -0.0458, 'grad_norm': 2.6527109146118164, 'learning_rate': 8.666666666666668e-06, 'num_tokens': 864646.0, 'completions/mean_length': 219.0, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 43.20000076293945, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.7206135988235474, 'rewards/meter/std': 0.34103232622146606, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.8617742657661438, 'rewards/lexical_plausibility/std': 0.12996774911880493, 'rewards/judge_quality/mean': 0.2837499976158142, 'rewards/judge_quality/std': 0.29271817207336426, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.20175716280937195, 'rewards/total_composite/std': 0.2930894196033478, 'reward': 0.20175716280937195, 'reward_std': 0.29308944940567017, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22974267601966858, 'sampling/sampling_logp_difference/max': 1.2344341278076172, 'sampling/importance_sampling_ratio/min': 0.2909993827342987, 'sampling/importance_sampling_ratio/mean': 1.0705015659332275, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3769593685865402, 'clip_ratio/low_mean': 0.07727272901684046, 'clip_ratio/low_min': 0.07727272901684046, 'clip_ratio/high_mean': 0.05633075162768364, 'clip_ratio/high_max': 0.05633075162768364, 'clip_ratio/region_mean': 0.1336034806445241, 'reward_total_mean': 0.20175716280937195, 'reward_meter_mean': 0.7206135988235474, 'reward_meter_std': 0.34103232622146606, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.8617742657661438, 'reward_lexical_plausibility_std': 0.12996774911880493, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.2837499976158142, 'reward_judge_quality_std': 0.29271817207336426, 'reward_total_composite_mean': 0.20175716280937195, 'reward_total_composite_std': 0.2930894196033478, 'epoch': 0.02}

 13%|█▎        | 441/3300 [1:39:59<11:49:57, 14.90s/it]INFO 04-12 12:26:11 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 442/3300 [1:40:14<11:49:37, 14.90s/it]2026-04-12 12:26:26,211 | INFO | train_grpo_train | metrics_logged mode=train step=442

                                                       
{'loss': -0.0639, 'grad_norm': 5.3352203369140625, 'learning_rate': 8.663636363636363e-06, 'num_tokens': 866691.0, 'completions/mean_length': 157.625, 'completions/min_length': 91.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 107.00000762939453, 'completions/min_terminated_length': 91.0, 'completions/max_terminated_length': 122.0, 'rewards/meter/mean': 0.5009660720825195, 'rewards/meter/std': 0.30664175748825073, 'rewards/count_adherence/mean': 0.78125, 'rewards/count_adherence/std': 0.24775780737400055, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9654681086540222, 'rewards/lexical_plausibility/std': 0.09767089784145355, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.1922609806060791, 'rewards/repeat_penalty/mean': 0.7729740738868713, 'rewards/repeat_penalty/std': 0.31112799048423767, 'rewards/near_duplicate_penalty/mean': 0.9449348449707031, 'rewards/near_duplicate_penalty/std': 0.07965335249900818, 'rewards/opening_diversity/mean': 0.9354166984558105, 'rewards/opening_diversity/std': 0.09776775538921356, 'rewards/distinct_2/mean': 0.8676998615264893, 'rewards/distinct_2/std': 0.2325456142425537, 'rewards/total_composite/mean': 0.12077479064464569, 'rewards/total_composite/std': 0.10433389246463776, 'reward': 0.12077479064464569, 'reward_std': 0.10433388501405716, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1649058312177658, 'sampling/sampling_logp_difference/max': 1.564957618713379, 'sampling/importance_sampling_ratio/min': 0.20909686386585236, 'sampling/importance_sampling_ratio/mean': 1.0242584943771362, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.209015041589737, 'clip_ratio/low_mean': 0.09367187134921551, 'clip_ratio/low_min': 0.09367187134921551, 'clip_ratio/high_mean': 0.03636094927787781, 'clip_ratio/high_max': 0.03636094927787781, 'clip_ratio/region_mean': 0.13003282062709332, 'reward_total_mean': 0.12077479064464569, 'reward_meter_mean': 0.5009660720825195, 'reward_meter_std': 0.30664175748825073, 'reward_count_adherence_mean': 0.78125, 'reward_count_adherence_std': 0.24775780737400055, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9654681086540222, 'reward_lexical_plausibility_std': 0.09767089784145355, 'reward_repeat_penalty_mean': 0.7729740738868713, 'reward_repeat_penalty_std': 0.31112799048423767, 'reward_near_duplicate_penalty_mean': 0.9449348449707031, 'reward_near_duplicate_penalty_std': 0.07965335249900818, 'reward_opening_diversity_mean': 0.9354166984558105, 'reward_opening_diversity_std': 0.09776775538921356, 'reward_distinct_2_mean': 0.8676998615264893, 'reward_distinct_2_std': 0.2325456142425537, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.1922609806060791, 'reward_total_composite_mean': 0.12077479064464569, 'reward_total_composite_std': 0.10433389246463776, 'epoch': 0.02}

 13%|█▎        | 442/3300 [1:40:14<11:49:37, 14.90s/it]INFO 04-12 12:26:26 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 443/3300 [1:40:28<11:42:12, 14.75s/it]2026-04-12 12:26:40,599 | INFO | train_grpo_train | metrics_logged mode=train step=443

                                                       
{'loss': -0.2596, 'grad_norm': 1.8034101724624634, 'learning_rate': 8.660606060606062e-06, 'num_tokens': 869181.0, 'completions/mean_length': 356.25, 'completions/min_length': 160.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 200.5, 'completions/min_terminated_length': 160.0, 'completions/max_terminated_length': 238.0, 'rewards/meter/mean': 0.6916443109512329, 'rewards/meter/std': 0.43410831689834595, 'rewards/count_adherence/mean': 0.5535714626312256, 'rewards/count_adherence/std': 0.3095892369747162, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.9266760349273682, 'rewards/lexical_plausibility/std': 0.10037942975759506, 'rewards/judge_quality/mean': 0.10000000149011612, 'rewards/judge_quality/std': 0.053452253341674805, 'rewards/repeat_penalty/mean': 0.47961631417274475, 'rewards/repeat_penalty/std': 0.47854146361351013, 'rewards/near_duplicate_penalty/mean': 0.8452885150909424, 'rewards/near_duplicate_penalty/std': 0.1606728583574295, 'rewards/opening_diversity/mean': 0.9388021230697632, 'rewards/opening_diversity/std': 0.0955413281917572, 'rewards/distinct_2/mean': 0.5390807390213013, 'rewards/distinct_2/std': 0.4962080419063568, 'rewards/total_composite/mean': 0.05557997524738312, 'rewards/total_composite/std': 0.05952258035540581, 'reward': 0.05557997524738312, 'reward_std': 0.05952257663011551, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.077547587454319, 'sampling/sampling_logp_difference/max': 2.1262025833129883, 'sampling/importance_sampling_ratio/min': 0.11928942799568176, 'sampling/importance_sampling_ratio/mean': 1.0103402137756348, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.36730458214879036, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.03248908510431647, 'clip_ratio/high_max': 0.03248908510431647, 'clip_ratio/region_mean': 0.03248908510431647, 'reward_total_mean': 0.05557997524738312, 'reward_meter_mean': 0.6916443109512329, 'reward_meter_std': 0.43410831689834595, 'reward_count_adherence_mean': 0.5535714626312256, 'reward_count_adherence_std': 0.3095892369747162, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.9266760349273682, 'reward_lexical_plausibility_std': 0.10037942975759506, 'reward_repeat_penalty_mean': 0.47961631417274475, 'reward_repeat_penalty_std': 0.47854146361351013, 'reward_near_duplicate_penalty_mean': 0.8452885150909424, 'reward_near_duplicate_penalty_std': 0.1606728583574295, 'reward_opening_diversity_mean': 0.9388021230697632, 'reward_opening_diversity_std': 0.0955413281917572, 'reward_distinct_2_mean': 0.5390807390213013, 'reward_distinct_2_std': 0.4962080419063568, 'reward_judge_quality_mean': 0.10000000149011612, 'reward_judge_quality_std': 0.053452253341674805, 'reward_total_composite_mean': 0.05557997524738312, 'reward_total_composite_std': 0.05952258035540581, 'epoch': 0.02}

 13%|█▎        | 443/3300 [1:40:28<11:42:12, 14.75s/it]INFO 04-12 12:26:40 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 444/3300 [1:40:43<11:47:44, 14.87s/it]2026-04-12 12:26:55,749 | INFO | train_grpo_train | metrics_logged mode=train step=444

                                                       
{'loss': -0.0573, 'grad_norm': 4.504018306732178, 'learning_rate': 8.657575757575758e-06, 'num_tokens': 871493.0, 'completions/mean_length': 179.0, 'completions/min_length': 108.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 131.42857360839844, 'completions/min_terminated_length': 108.0, 'completions/max_terminated_length': 183.0, 'rewards/meter/mean': 0.7618038654327393, 'rewards/meter/std': 0.3433324694633484, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.10690449178218842, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.974048376083374, 'rewards/lexical_plausibility/std': 0.07340221107006073, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.14078859984874725, 'rewards/repeat_penalty/mean': 0.6010661125183105, 'rewards/repeat_penalty/std': 0.3114267885684967, 'rewards/near_duplicate_penalty/mean': 0.8809487223625183, 'rewards/near_duplicate_penalty/std': 0.16103245317935944, 'rewards/opening_diversity/mean': 0.9261363744735718, 'rewards/opening_diversity/std': 0.20891793072223663, 'rewards/distinct_2/mean': 0.6647645235061646, 'rewards/distinct_2/std': 0.31705909967422485, 'rewards/total_composite/mean': 0.19647744297981262, 'rewards/total_composite/std': 0.13350331783294678, 'reward': 0.19647744297981262, 'reward_std': 0.13350333273410797, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1287020742893219, 'sampling/sampling_logp_difference/max': 2.543274164199829, 'sampling/importance_sampling_ratio/min': 0.07860860228538513, 'sampling/importance_sampling_ratio/mean': 1.0116922855377197, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7886410430073738, 'clip_ratio/low_mean': 0.01476200157776475, 'clip_ratio/low_min': 0.01476200157776475, 'clip_ratio/high_mean': 0.09079050365835428, 'clip_ratio/high_max': 0.09079050365835428, 'clip_ratio/region_mean': 0.10555250523611903, 'reward_total_mean': 0.19647744297981262, 'reward_meter_mean': 0.7618038654327393, 'reward_meter_std': 0.3433324694633484, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.10690449178218842, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.974048376083374, 'reward_lexical_plausibility_std': 0.07340221107006073, 'reward_repeat_penalty_mean': 0.6010661125183105, 'reward_repeat_penalty_std': 0.3114267885684967, 'reward_near_duplicate_penalty_mean': 0.8809487223625183, 'reward_near_duplicate_penalty_std': 0.16103245317935944, 'reward_opening_diversity_mean': 0.9261363744735718, 'reward_opening_diversity_std': 0.20891793072223663, 'reward_distinct_2_mean': 0.6647645235061646, 'reward_distinct_2_std': 0.31705909967422485, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.14078859984874725, 'reward_total_composite_mean': 0.19647744297981262, 'reward_total_composite_std': 0.13350331783294678, 'epoch': 0.02}

 13%|█▎        | 444/3300 [1:40:43<11:47:44, 14.87s/it]INFO 04-12 12:26:55 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 445/3300 [1:40:57<11:33:59, 14.58s/it]2026-04-12 12:27:09,675 | INFO | train_grpo_train | metrics_logged mode=train step=445

                                                       
{'loss': -0.159, 'grad_norm': 2.706441879272461, 'learning_rate': 8.654545454545455e-06, 'num_tokens': 873899.0, 'completions/mean_length': 190.75, 'completions/min_length': 129.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 144.85714721679688, 'completions/min_terminated_length': 129.0, 'completions/max_terminated_length': 156.0, 'rewards/meter/mean': 0.9669731855392456, 'rewards/meter/std': 0.028420019894838333, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.2357022613286972, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9743486046791077, 'rewards/lexical_plausibility/std': 0.07255306839942932, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.12464234977960587, 'rewards/repeat_penalty/mean': 0.22924906015396118, 'rewards/repeat_penalty/std': 0.3353690207004547, 'rewards/near_duplicate_penalty/mean': 0.7837797403335571, 'rewards/near_duplicate_penalty/std': 0.14415282011032104, 'rewards/opening_diversity/mean': 0.8776785135269165, 'rewards/opening_diversity/std': 0.10486959666013718, 'rewards/distinct_2/mean': 0.2895064651966095, 'rewards/distinct_2/std': 0.3314993977546692, 'rewards/total_composite/mean': 0.20554478466510773, 'rewards/total_composite/std': 0.11008559167385101, 'reward': 0.20554478466510773, 'reward_std': 0.11008559167385101, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0865015834569931, 'sampling/sampling_logp_difference/max': 1.572070598602295, 'sampling/importance_sampling_ratio/min': 0.20761485397815704, 'sampling/importance_sampling_ratio/mean': 1.0027402639389038, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5501820705831051, 'clip_ratio/low_mean': 0.015247253235429525, 'clip_ratio/low_min': 0.015247253235429525, 'clip_ratio/high_mean': 0.06887097377330065, 'clip_ratio/high_max': 0.06887097377330065, 'clip_ratio/region_mean': 0.08411822700873017, 'reward_total_mean': 0.20554478466510773, 'reward_meter_mean': 0.9669731855392456, 'reward_meter_std': 0.028420019894838333, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.2357022613286972, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9743486046791077, 'reward_lexical_plausibility_std': 0.07255306839942932, 'reward_repeat_penalty_mean': 0.22924906015396118, 'reward_repeat_penalty_std': 0.3353690207004547, 'reward_near_duplicate_penalty_mean': 0.7837797403335571, 'reward_near_duplicate_penalty_std': 0.14415282011032104, 'reward_opening_diversity_mean': 0.8776785135269165, 'reward_opening_diversity_std': 0.10486959666013718, 'reward_distinct_2_mean': 0.2895064651966095, 'reward_distinct_2_std': 0.3314993977546692, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.12464234977960587, 'reward_total_composite_mean': 0.20554478466510773, 'reward_total_composite_std': 0.11008559167385101, 'epoch': 0.02}

 13%|█▎        | 445/3300 [1:40:57<11:33:59, 14.58s/it]INFO 04-12 12:27:09 [block_pool.py:378] Successfully reset prefix cache

 14%|█▎        | 446/3300 [1:41:13<11:46:24, 14.85s/it]2026-04-12 12:27:25,146 | INFO | train_grpo_train | metrics_logged mode=train step=446

                                                       
{'loss': -0.0779, 'grad_norm': 5.328366756439209, 'learning_rate': 8.651515151515152e-06, 'num_tokens': 876208.0, 'completions/mean_length': 163.625, 'completions/min_length': 93.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 113.85714721679688, 'completions/min_terminated_length': 93.0, 'completions/max_terminated_length': 141.0, 'rewards/meter/mean': 0.8344897031784058, 'rewards/meter/std': 0.24713453650474548, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9974895715713501, 'rewards/lexical_plausibility/std': 0.007100627291947603, 'rewards/judge_quality/mean': 0.47999995946884155, 'rewards/judge_quality/std': 0.29871153831481934, 'rewards/repeat_penalty/mean': 0.9614061117172241, 'rewards/repeat_penalty/std': 0.08976607024669647, 'rewards/near_duplicate_penalty/mean': 0.9900728464126587, 'rewards/near_duplicate_penalty/std': 0.009938557632267475, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9704141616821289, 'rewards/distinct_2/std': 0.08368128538131714, 'rewards/total_composite/mean': 0.316730797290802, 'rewards/total_composite/std': 0.26133811473846436, 'reward': 0.316730797290802, 'reward_std': 0.26133811473846436, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1635666936635971, 'sampling/sampling_logp_difference/max': 1.8835573196411133, 'sampling/importance_sampling_ratio/min': 0.152048259973526, 'sampling/importance_sampling_ratio/mean': 1.0204392671585083, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2078724652528763, 'clip_ratio/low_mean': 0.09674250427633524, 'clip_ratio/low_min': 0.09674250427633524, 'clip_ratio/high_mean': 0.035340793430805206, 'clip_ratio/high_max': 0.035340793430805206, 'clip_ratio/region_mean': 0.13208329770714045, 'reward_total_mean': 0.316730797290802, 'reward_meter_mean': 0.8344897031784058, 'reward_meter_std': 0.24713453650474548, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9974895715713501, 'reward_lexical_plausibility_std': 0.007100627291947603, 'reward_repeat_penalty_mean': 0.9614061117172241, 'reward_repeat_penalty_std': 0.08976607024669647, 'reward_near_duplicate_penalty_mean': 0.9900728464126587, 'reward_near_duplicate_penalty_std': 0.009938557632267475, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9704141616821289, 'reward_distinct_2_std': 0.08368128538131714, 'reward_judge_quality_mean': 0.47999995946884155, 'reward_judge_quality_std': 0.29871153831481934, 'reward_total_composite_mean': 0.316730797290802, 'reward_total_composite_std': 0.26133811473846436, 'epoch': 0.02}

 14%|█▎        | 446/3300 [1:41:13<11:46:24, 14.85s/it]INFO 04-12 12:27:25 [block_pool.py:378] Successfully reset prefix cache

 14%|█▎        | 447/3300 [1:41:23<10:47:33, 13.62s/it]2026-04-12 12:27:35,893 | INFO | train_grpo_train | metrics_logged mode=train step=447

                                                       
{'loss': 0.0349, 'grad_norm': 11.789586067199707, 'learning_rate': 8.64848484848485e-06, 'num_tokens': 878217.0, 'completions/mean_length': 60.125, 'completions/min_length': 54.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.125, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.7874279022216797, 'rewards/meter/std': 0.31900274753570557, '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/lexical_plausibility/mean': 0.9917200803756714, 'rewards/lexical_plausibility/std': 0.015595059841871262, 'rewards/judge_quality/mean': 0.4624999761581421, 'rewards/judge_quality/std': 0.20310096442699432, 'rewards/repeat_penalty/mean': 0.90325528383255, 'rewards/repeat_penalty/std': 0.12485902011394501, 'rewards/near_duplicate_penalty/mean': 0.9607804417610168, 'rewards/near_duplicate_penalty/std': 0.043191585689783096, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9676003456115723, 'rewards/distinct_2/std': 0.04872959107160568, 'rewards/total_composite/mean': 0.37079668045043945, 'rewards/total_composite/std': 0.2684282660484314, 'reward': 0.37079668045043945, 'reward_std': 0.268428236246109, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16090291738510132, 'sampling/sampling_logp_difference/max': 1.4209400415420532, 'sampling/importance_sampling_ratio/min': 0.3209053575992584, 'sampling/importance_sampling_ratio/mean': 1.0293289422988892, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2205371633172035, 'clip_ratio/low_mean': 0.08512650988996029, 'clip_ratio/low_min': 0.08512650988996029, 'clip_ratio/high_mean': 0.06675884500145912, 'clip_ratio/high_max': 0.06675884500145912, 'clip_ratio/region_mean': 0.1518853548914194, 'reward_total_mean': 0.37079668045043945, 'reward_meter_mean': 0.7874279022216797, 'reward_meter_std': 0.31900274753570557, '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_lexical_plausibility_mean': 0.9917200803756714, 'reward_lexical_plausibility_std': 0.015595059841871262, 'reward_repeat_penalty_mean': 0.90325528383255, 'reward_repeat_penalty_std': 0.12485902011394501, 'reward_near_duplicate_penalty_mean': 0.9607804417610168, 'reward_near_duplicate_penalty_std': 0.043191585689783096, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9676003456115723, 'reward_distinct_2_std': 0.04872959107160568, 'reward_judge_quality_mean': 0.4624999761581421, 'reward_judge_quality_std': 0.20310096442699432, 'reward_total_composite_mean': 0.37079668045043945, 'reward_total_composite_std': 0.2684282660484314, 'epoch': 0.02}

 14%|█▎        | 447/3300 [1:41:23<10:47:33, 13.62s/it]INFO 04-12 12:27:36 [block_pool.py:378] Successfully reset prefix cache

 14%|█▎        | 448/3300 [1:41:38<11:04:59, 13.99s/it]2026-04-12 12:27:50,747 | INFO | train_grpo_train | metrics_logged mode=train step=448

                                                       
{'loss': -0.0535, 'grad_norm': 4.804816722869873, 'learning_rate': 8.645454545454545e-06, 'num_tokens': 879862.0, 'completions/mean_length': 105.625, 'completions/min_length': 43.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 47.57143020629883, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.5961958765983582, 'rewards/meter/std': 0.3931199014186859, '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/lexical_plausibility/mean': 0.9792802929878235, 'rewards/lexical_plausibility/std': 0.0586041621863842, 'rewards/judge_quality/mean': 0.5375000238418579, 'rewards/judge_quality/std': 0.3347814083099365, 'rewards/repeat_penalty/mean': 0.9795229434967041, 'rewards/repeat_penalty/std': 0.020922917872667313, 'rewards/near_duplicate_penalty/mean': 0.9795229434967041, 'rewards/near_duplicate_penalty/std': 0.020922917872667313, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.36088132858276367, 'rewards/total_composite/std': 0.33048778772354126, 'reward': 0.36088132858276367, 'reward_std': 0.33048778772354126, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1440524011850357, 'sampling/sampling_logp_difference/max': 1.861936092376709, 'sampling/importance_sampling_ratio/min': 0.1553715318441391, 'sampling/importance_sampling_ratio/mean': 1.0400636196136475, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9928457960486412, 'clip_ratio/low_mean': 0.08905484667047858, 'clip_ratio/low_min': 0.08905484667047858, 'clip_ratio/high_mean': 0.046394798904657364, 'clip_ratio/high_max': 0.046394798904657364, 'clip_ratio/region_mean': 0.13544964557513595, 'reward_total_mean': 0.36088132858276367, 'reward_meter_mean': 0.5961958765983582, 'reward_meter_std': 0.3931199014186859, '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_lexical_plausibility_mean': 0.9792802929878235, 'reward_lexical_plausibility_std': 0.0586041621863842, 'reward_repeat_penalty_mean': 0.9795229434967041, 'reward_repeat_penalty_std': 0.020922917872667313, 'reward_near_duplicate_penalty_mean': 0.9795229434967041, 'reward_near_duplicate_penalty_std': 0.020922917872667313, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5375000238418579, 'reward_judge_quality_std': 0.3347814083099365, 'reward_total_composite_mean': 0.36088132858276367, 'reward_total_composite_std': 0.33048778772354126, 'epoch': 0.02}

 14%|█▎        | 448/3300 [1:41:38<11:04:59, 13.99s/it]INFO 04-12 12:27:51 [block_pool.py:378] Successfully reset prefix cache

 14%|█▎        | 449/3300 [1:41:53<11:16:58, 14.25s/it]2026-04-12 12:28:05,593 | INFO | train_grpo_train | metrics_logged mode=train step=449

                                                       
{'loss': -0.3189, 'grad_norm': 2.427307605743408, 'learning_rate': 8.642424242424242e-06, 'num_tokens': 882961.0, 'completions/mean_length': 327.375, 'completions/min_length': 233.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 265.8333435058594, 'completions/min_terminated_length': 233.0, 'completions/max_terminated_length': 340.0, 'rewards/meter/mean': 0.8265117406845093, 'rewards/meter/std': 0.33677634596824646, 'rewards/count_adherence/mean': 0.7045454978942871, 'rewards/count_adherence/std': 0.33224931359291077, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9413131475448608, 'rewards/lexical_plausibility/std': 0.11571116000413895, 'rewards/judge_quality/mean': 0.13750000298023224, 'rewards/judge_quality/std': 0.06408699601888657, 'rewards/repeat_penalty/mean': 0.21820887923240662, 'rewards/repeat_penalty/std': 0.4093443751335144, 'rewards/near_duplicate_penalty/mean': 0.6777788400650024, 'rewards/near_duplicate_penalty/std': 0.2975876033306122, 'rewards/opening_diversity/mean': 0.7090300917625427, 'rewards/opening_diversity/std': 0.33719363808631897, 'rewards/distinct_2/mean': 0.25, 'rewards/distinct_2/std': 0.4629100561141968, 'rewards/total_composite/mean': 0.10369314253330231, 'rewards/total_composite/std': 0.06705006211996078, 'reward': 0.10369314253330231, 'reward_std': 0.06705006211996078, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05021149665117264, 'sampling/sampling_logp_difference/max': 2.2314655780792236, 'sampling/importance_sampling_ratio/min': 0.10737095773220062, 'sampling/importance_sampling_ratio/mean': 1.0050650835037231, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2698381785303354, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.03363513865042478, 'clip_ratio/high_max': 0.03363513865042478, 'clip_ratio/region_mean': 0.03363513865042478, 'reward_total_mean': 0.10369314253330231, 'reward_meter_mean': 0.8265117406845093, 'reward_meter_std': 0.33677634596824646, 'reward_count_adherence_mean': 0.7045454978942871, 'reward_count_adherence_std': 0.33224931359291077, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9413131475448608, 'reward_lexical_plausibility_std': 0.11571116000413895, 'reward_repeat_penalty_mean': 0.21820887923240662, 'reward_repeat_penalty_std': 0.4093443751335144, 'reward_near_duplicate_penalty_mean': 0.6777788400650024, 'reward_near_duplicate_penalty_std': 0.2975876033306122, 'reward_opening_diversity_mean': 0.7090300917625427, 'reward_opening_diversity_std': 0.33719363808631897, 'reward_distinct_2_mean': 0.25, 'reward_distinct_2_std': 0.4629100561141968, 'reward_judge_quality_mean': 0.13750000298023224, 'reward_judge_quality_std': 0.06408699601888657, 'reward_total_composite_mean': 0.10369314253330231, 'reward_total_composite_std': 0.06705006211996078, 'epoch': 0.02}

 14%|█▎        | 449/3300 [1:41:53<11:16:58, 14.25s/it]INFO 04-12 12:28:05 [block_pool.py:378] Successfully reset prefix cache

 14%|█▎        | 450/3300 [1:42:08<11:28:12, 14.49s/it]2026-04-12 12:28:20,649 | INFO | train_grpo_train | metrics_logged mode=train step=450

                                                       
{'loss': -0.2385, 'grad_norm': 2.4964518547058105, 'learning_rate': 8.63939393939394e-06, 'num_tokens': 887055.0, 'completions/mean_length': 360.75, 'completions/min_length': 326.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 339.14288330078125, 'completions/min_terminated_length': 326.0, 'completions/max_terminated_length': 352.0, 'rewards/meter/mean': 0.8051550388336182, 'rewards/meter/std': 0.3649185001850128, 'rewards/count_adherence/mean': 0.6499999761581421, 'rewards/count_adherence/std': 0.22677867114543915, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9714911580085754, 'rewards/lexical_plausibility/std': 0.08063514530658722, 'rewards/judge_quality/mean': 0.13750000298023224, 'rewards/judge_quality/std': 0.0353553406894207, 'rewards/repeat_penalty/mean': 0.2087945193052292, 'rewards/repeat_penalty/std': 0.35941508412361145, 'rewards/near_duplicate_penalty/mean': 0.7662945985794067, 'rewards/near_duplicate_penalty/std': 0.16975484788417816, 'rewards/opening_diversity/mean': 0.8149038553237915, 'rewards/opening_diversity/std': 0.2093890756368637, 'rewards/distinct_2/mean': 0.25672823190689087, 'rewards/distinct_2/std': 0.4299718141555786, 'rewards/total_composite/mean': 0.088001549243927, 'rewards/total_composite/std': 0.04168659821152687, 'reward': 0.088001549243927, 'reward_std': 0.04168659821152687, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07258587330579758, 'sampling/sampling_logp_difference/max': 2.150117874145508, 'sampling/importance_sampling_ratio/min': 0.11647042632102966, 'sampling/importance_sampling_ratio/mean': 1.0100191831588745, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7186020761728287, 'clip_ratio/low_mean': 0.019396550953388214, 'clip_ratio/low_min': 0.019396550953388214, 'clip_ratio/high_mean': 0.040890860138460994, 'clip_ratio/high_max': 0.040890860138460994, 'clip_ratio/region_mean': 0.06028741109184921, 'reward_total_mean': 0.088001549243927, 'reward_meter_mean': 0.8051550388336182, 'reward_meter_std': 0.3649185001850128, 'reward_count_adherence_mean': 0.6499999761581421, 'reward_count_adherence_std': 0.22677867114543915, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9714911580085754, 'reward_lexical_plausibility_std': 0.08063514530658722, 'reward_repeat_penalty_mean': 0.2087945193052292, 'reward_repeat_penalty_std': 0.35941508412361145, 'reward_near_duplicate_penalty_mean': 0.7662945985794067, 'reward_near_duplicate_penalty_std': 0.16975484788417816, 'reward_opening_diversity_mean': 0.8149038553237915, 'reward_opening_diversity_std': 0.2093890756368637, 'reward_distinct_2_mean': 0.25672823190689087, 'reward_distinct_2_std': 0.4299718141555786, 'reward_judge_quality_mean': 0.13750000298023224, 'reward_judge_quality_std': 0.0353553406894207, 'reward_total_composite_mean': 0.088001549243927, 'reward_total_composite_std': 0.04168659821152687, 'epoch': 0.02}

 14%|█▎        | 450/3300 [1:42:08<11:28:12, 14.49s/it]INFO 04-12 12:28:20 [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 [00:10<00:59,  5.44s/it]

 23%|██▎       | 3/13 [00:18<01:01,  6.16s/it]

 31%|███       | 4/13 [00:29<01:11,  7.97s/it]

 38%|███▊      | 5/13 [00:39<01:11,  8.94s/it]

 46%|████▌     | 6/13 [00:51<01:09,  9.98s/it]

 54%|█████▍    | 7/13 [01:02<01:01, 10.26s/it]

 62%|██████▏   | 8/13 [01:14<00:54, 10.82s/it]

 69%|██████▉   | 9/13 [01:26<00:44, 11.16s/it]

 77%|███████▋  | 10/13 [01:39<00:34, 11.63s/it]

 85%|████████▍ | 11/13 [01:51<00:23, 11.71s/it]

 92%|█████████▏| 12/13 [02:04<00:12, 12.24s/it]

100%|██████████| 13/13 [02:17<00:00, 12.26s/it]2026-04-12 12:30:52,096 | INFO | train_grpo_train | metrics_logged mode=eval step=450

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 151.4077, 'eval_samples_per_second': 0.687, 'eval_steps_per_second': 0.086, 'eval_num_tokens': 887055.0, 'eval_completions/mean_length': 210.3653846153846, 'eval_completions/min_length': 46.69230769230769, 'eval_completions/max_length': 458.15384615384613, 'eval_completions/clipped_ratio': 0.1346153846153846, 'eval_completions/mean_terminated_length': 162.63773228571966, 'eval_completions/min_terminated_length': 46.69230769230769, 'eval_completions/max_terminated_length': 336.0769230769231, 'eval_rewards/meter/mean': 0.5266091227531433, 'eval_rewards/meter/std': 0.38933635445741505, 'eval_rewards/count_adherence/mean': 0.8129428441707904, 'eval_rewards/count_adherence/std': 0.21615326289947218, 'eval_rewards/arabic_clean/mean': 0.875, 'eval_rewards/arabic_clean/std': 0.25542253255844116, 'eval_rewards/lexical_plausibility/mean': 0.9771601878679715, 'eval_rewards/lexical_plausibility/std': 0.03900533656661327, 'eval_rewards/judge_quality/mean': 0.3425961480690883, 'eval_rewards/judge_quality/std': 0.19757812240949044, 'eval_rewards/repeat_penalty/mean': 0.6584883446876819, 'eval_rewards/repeat_penalty/std': 0.3737167280453902, 'eval_rewards/near_duplicate_penalty/mean': 0.905442462517665, 'eval_rewards/near_duplicate_penalty/std': 0.13739756408792275, 'eval_rewards/opening_diversity/mean': 0.9148693084716797, 'eval_rewards/opening_diversity/std': 0.15748956054449081, 'eval_rewards/distinct_2/mean': 0.7157056377484248, 'eval_rewards/distinct_2/std': 0.3620255205493707, 'eval_rewards/total_composite/mean': 0.15759526537014887, 'eval_rewards/total_composite/std': 0.15852333089480034, 'eval_reward': 0.15759526537014887, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.06912292196200444, 'eval_sampling/sampling_logp_difference/max': 1.1673028285686786, 'eval_sampling/importance_sampling_ratio/min': 0.31652708466236407, 'eval_sampling/importance_sampling_ratio/mean': 1.0160807646237886, 'eval_sampling/importance_sampling_ratio/max': 1.53141293158898, 'eval_entropy': 0.8395213943261367, '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': 0.15759526537014887, 'eval_reward_meter_mean': 0.5266091227531433, 'eval_reward_meter_std': 0.38933635445741505, 'eval_reward_count_adherence_mean': 0.8129428441707904, 'eval_reward_count_adherence_std': 0.21615326289947218, 'eval_reward_arabic_clean_mean': 0.875, 'eval_reward_arabic_clean_std': 0.25542253255844116, 'eval_reward_lexical_plausibility_mean': 0.9771601878679715, 'eval_reward_lexical_plausibility_std': 0.03900533656661327, 'eval_reward_repeat_penalty_mean': 0.6584883446876819, 'eval_reward_repeat_penalty_std': 0.3737167280453902, 'eval_reward_near_duplicate_penalty_mean': 0.905442462517665, 'eval_reward_near_duplicate_penalty_std': 0.13739756408792275, 'eval_reward_opening_diversity_mean': 0.9148693084716797, 'eval_reward_opening_diversity_std': 0.15748956054449081, 'eval_reward_distinct_2_mean': 0.7157056377484248, 'eval_reward_distinct_2_std': 0.3620255205493707, 'eval_reward_judge_quality_mean': 0.3425961480690883, 'eval_reward_judge_quality_std': 0.19757812240949044, 'eval_reward_total_composite_mean': 0.15759526537014887, 'eval_reward_total_composite_std': 0.15852333089480034, 'epoch': 0.02}

 14%|█▎        | 450/3300 [1:44:40<11:28:12, 14.49s/it]

100%|██████████| 13/13 [02:17<00:00, 12.26s/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(

 14%|█▎        | 451/3300 [1:44:56<47:46:54, 60.38s/it]2026-04-12 12:31:08,099 | INFO | train_grpo_train | metrics_logged mode=train step=451

                                                       
{'loss': -0.0706, 'grad_norm': 2.170783042907715, 'learning_rate': 8.636363636363637e-06, 'num_tokens': 888626.0, 'completions/mean_length': 221.375, 'completions/min_length': 44.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 47.0, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.6213947534561157, 'rewards/meter/std': 0.4637322723865509, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.8898974657058716, 'rewards/lexical_plausibility/std': 0.15224117040634155, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.27998724579811096, 'rewards/repeat_penalty/mean': 0.9076778888702393, 'rewards/repeat_penalty/std': 0.16397406160831451, 'rewards/near_duplicate_penalty/mean': 0.9776195883750916, 'rewards/near_duplicate_penalty/std': 0.04622026905417442, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9825174808502197, 'rewards/distinct_2/std': 0.049448031932115555, 'rewards/total_composite/mean': 0.25828826427459717, 'rewards/total_composite/std': 0.2804103493690491, 'reward': 0.25828826427459717, 'reward_std': 0.28041037917137146, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18175174295902252, 'sampling/sampling_logp_difference/max': 1.9918127059936523, 'sampling/importance_sampling_ratio/min': 0.13644786179065704, 'sampling/importance_sampling_ratio/mean': 1.00529944896698, 'sampling/importance_sampling_ratio/max': 1.8992918729782104, 'entropy': 0.8205755650997162, 'clip_ratio/low_mean': 0.019886363297700882, 'clip_ratio/low_min': 0.019886363297700882, 'clip_ratio/high_mean': 0.08158301655203104, 'clip_ratio/high_max': 0.08158301655203104, 'clip_ratio/region_mean': 0.10146937984973192, 'reward_total_mean': 0.25828826427459717, 'reward_meter_mean': 0.6213947534561157, 'reward_meter_std': 0.4637322723865509, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.8898974657058716, 'reward_lexical_plausibility_std': 0.15224117040634155, 'reward_repeat_penalty_mean': 0.9076778888702393, 'reward_repeat_penalty_std': 0.16397406160831451, 'reward_near_duplicate_penalty_mean': 0.9776195883750916, 'reward_near_duplicate_penalty_std': 0.04622026905417442, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9825174808502197, 'reward_distinct_2_std': 0.049448031932115555, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.27998724579811096, 'reward_total_composite_mean': 0.25828826427459717, 'reward_total_composite_std': 0.2804103493690491, 'epoch': 0.02}

 14%|█▎        | 451/3300 [1:44:56<47:46:54, 60.38s/it]INFO 04-12 12:31:08 [block_pool.py:378] Successfully reset prefix cache

 14%|█▎        | 452/3300 [1:45:05<35:46:58, 45.23s/it]2026-04-12 12:31:17,986 | INFO | train_grpo_train | metrics_logged mode=train step=452

                                                       
{'loss': 0.0491, 'grad_norm': 16.583702087402344, 'learning_rate': 8.633333333333334e-06, 'num_tokens': 890370.0, 'completions/mean_length': 50.0, 'completions/min_length': 43.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.0, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.8257185816764832, 'rewards/meter/std': 0.2584497332572937, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4962500035762787, 'rewards/judge_quality/std': 0.17476005852222443, 'rewards/repeat_penalty/mean': 0.9797744750976562, 'rewards/repeat_penalty/std': 0.0284157395362854, 'rewards/near_duplicate_penalty/mean': 0.9797744750976562, 'rewards/near_duplicate_penalty/std': 0.0284157395362854, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4004746079444885, 'rewards/total_composite/std': 0.14909245073795319, 'reward': 0.4004746079444885, 'reward_std': 0.149092435836792, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17566032707691193, 'sampling/sampling_logp_difference/max': 1.372208595275879, 'sampling/importance_sampling_ratio/min': 0.2535463571548462, 'sampling/importance_sampling_ratio/mean': 1.0329374074935913, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5124791264533997, 'clip_ratio/low_mean': 0.07602121867239475, 'clip_ratio/low_min': 0.07602121867239475, 'clip_ratio/high_mean': 0.07484584953635931, 'clip_ratio/high_max': 0.07484584953635931, 'clip_ratio/region_mean': 0.15086706820875406, 'reward_total_mean': 0.4004746079444885, 'reward_meter_mean': 0.8257185816764832, 'reward_meter_std': 0.2584497332572937, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9797744750976562, 'reward_repeat_penalty_std': 0.0284157395362854, 'reward_near_duplicate_penalty_mean': 0.9797744750976562, 'reward_near_duplicate_penalty_std': 0.0284157395362854, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4962500035762787, 'reward_judge_quality_std': 0.17476005852222443, 'reward_total_composite_mean': 0.4004746079444885, 'reward_total_composite_std': 0.14909245073795319, 'epoch': 0.02}

 14%|█▎        | 452/3300 [1:45:05<35:46:58, 45.23s/it]INFO 04-12 12:31:18 [block_pool.py:378] Successfully reset prefix cache

 14%|█▎        | 453/3300 [1:45:16<27:34:25, 34.87s/it]2026-04-12 12:31:28,677 | INFO | train_grpo_train | metrics_logged mode=train step=453

                                                       
{'loss': 0.13, 'grad_norm': 6.972647666931152, 'learning_rate': 8.630303030303032e-06, 'num_tokens': 892783.0, 'completions/mean_length': 121.625, 'completions/min_length': 99.0, 'completions/max_length': 156.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 121.625, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 156.0, 'rewards/meter/mean': 0.9071915149688721, 'rewards/meter/std': 0.09999934583902359, 'rewards/count_adherence/mean': 0.8500000238418579, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/repeat_penalty/mean': 0.4688087999820709, 'rewards/repeat_penalty/std': 0.36062806844711304, 'rewards/near_duplicate_penalty/mean': 0.8616430759429932, 'rewards/near_duplicate_penalty/std': 0.14222761988639832, 'rewards/opening_diversity/mean': 0.859375, 'rewards/opening_diversity/std': 0.302057683467865, 'rewards/distinct_2/mean': 0.5397911071777344, 'rewards/distinct_2/std': 0.3427259922027588, 'rewards/total_composite/mean': 0.2506418228149414, 'rewards/total_composite/std': 0.06561446934938431, 'reward': 0.2506418228149414, 'reward_std': 0.06561446934938431, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11622463166713715, 'sampling/sampling_logp_difference/max': 2.561680793762207, 'sampling/importance_sampling_ratio/min': 0.07717491686344147, 'sampling/importance_sampling_ratio/mean': 1.0179163217544556, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6076498217880726, 'clip_ratio/low_mean': 0.03519226820208132, 'clip_ratio/low_min': 0.03519226820208132, 'clip_ratio/high_mean': 0.061575477942824364, 'clip_ratio/high_max': 0.061575477942824364, 'clip_ratio/region_mean': 0.09676774614490569, 'reward_total_mean': 0.2506418228149414, 'reward_meter_mean': 0.9071915149688721, 'reward_meter_std': 0.09999934583902359, 'reward_count_adherence_mean': 0.8500000238418579, 'reward_count_adherence_std': 0.09258200973272324, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.4688087999820709, 'reward_repeat_penalty_std': 0.36062806844711304, 'reward_near_duplicate_penalty_mean': 0.8616430759429932, 'reward_near_duplicate_penalty_std': 0.14222761988639832, 'reward_opening_diversity_mean': 0.859375, 'reward_opening_diversity_std': 0.302057683467865, 'reward_distinct_2_mean': 0.5397911071777344, 'reward_distinct_2_std': 0.3427259922027588, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.2506418228149414, 'reward_total_composite_std': 0.06561446934938431, 'epoch': 0.02}

 14%|█▎        | 453/3300 [1:45:16<27:34:25, 34.87s/it]INFO 04-12 12:31:28 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 454/3300 [1:45:30<22:33:29, 28.53s/it]2026-04-12 12:31:42,428 | INFO | train_grpo_train | metrics_logged mode=train step=454

                                                       
{'loss': -0.0411, 'grad_norm': 1.3595945835113525, 'learning_rate': 8.627272727272727e-06, 'num_tokens': 894238.0, 'completions/mean_length': 139.875, 'completions/min_length': 14.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 15.833333969116211, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 18.0, 'rewards/meter/mean': 0.7350038290023804, 'rewards/meter/std': 0.41689085960388184, '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/lexical_plausibility/mean': 0.9627242088317871, 'rewards/lexical_plausibility/std': 0.0736796110868454, 'rewards/judge_quality/mean': 0.6312500238418579, 'rewards/judge_quality/std': 0.40912583470344543, 'rewards/repeat_penalty/mean': 0.5043357014656067, 'rewards/repeat_penalty/std': 0.2692793905735016, 'rewards/near_duplicate_penalty/mean': 0.6847714781761169, 'rewards/near_duplicate_penalty/std': 0.2506482005119324, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.8889120817184448, 'rewards/distinct_2/std': 0.13700076937675476, 'rewards/total_composite/mean': 0.5911006331443787, 'rewards/total_composite/std': 0.41112080216407776, 'reward': 0.5911006331443787, 'reward_std': 0.41112077236175537, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.139166459441185, 'sampling/sampling_logp_difference/max': 1.0971436500549316, 'sampling/importance_sampling_ratio/min': 0.33382323384284973, 'sampling/importance_sampling_ratio/mean': 0.9744467735290527, 'sampling/importance_sampling_ratio/max': 1.6231614351272583, 'entropy': 0.6575254499912262, 'clip_ratio/low_mean': 0.008928571827709675, 'clip_ratio/low_min': 0.008928571827709675, 'clip_ratio/high_mean': 0.0935763893648982, 'clip_ratio/high_max': 0.0935763893648982, 'clip_ratio/region_mean': 0.10250496119260788, 'reward_total_mean': 0.5911006331443787, 'reward_meter_mean': 0.7350038290023804, 'reward_meter_std': 0.41689085960388184, '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_lexical_plausibility_mean': 0.9627242088317871, 'reward_lexical_plausibility_std': 0.0736796110868454, 'reward_repeat_penalty_mean': 0.5043357014656067, 'reward_repeat_penalty_std': 0.2692793905735016, 'reward_near_duplicate_penalty_mean': 0.6847714781761169, 'reward_near_duplicate_penalty_std': 0.2506482005119324, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.8889120817184448, 'reward_distinct_2_std': 0.13700076937675476, 'reward_judge_quality_mean': 0.6312500238418579, 'reward_judge_quality_std': 0.40912583470344543, 'reward_total_composite_mean': 0.5911006331443787, 'reward_total_composite_std': 0.41112080216407776, 'epoch': 0.02}

 14%|█▍        | 454/3300 [1:45:30<22:33:29, 28.53s/it]INFO 04-12 12:31:42 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 455/3300 [1:45:43<18:59:32, 24.03s/it]2026-04-12 12:31:55,964 | INFO | train_grpo_train | metrics_logged mode=train step=455

                                                       
{'loss': -0.0013, 'grad_norm': 6.438784122467041, 'learning_rate': 8.624242424242424e-06, 'num_tokens': 895717.0, 'completions/mean_length': 101.875, 'completions/min_length': 36.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 43.28571701049805, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.8325039148330688, 'rewards/meter/std': 0.2588021159172058, '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/lexical_plausibility/mean': 0.9751418232917786, 'rewards/lexical_plausibility/std': 0.070309579372406, 'rewards/judge_quality/mean': 0.4675000309944153, 'rewards/judge_quality/std': 0.3998839259147644, 'rewards/repeat_penalty/mean': 0.9999389052391052, 'rewards/repeat_penalty/std': 0.00017280207248404622, 'rewards/near_duplicate_penalty/mean': 0.9999389052391052, 'rewards/near_duplicate_penalty/std': 0.00017280207248404622, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3942283093929291, 'rewards/total_composite/std': 0.35926198959350586, 'reward': 0.3942283093929291, 'reward_std': 0.35926195979118347, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17224042117595673, 'sampling/sampling_logp_difference/max': 1.1990406513214111, 'sampling/importance_sampling_ratio/min': 0.30148330330848694, 'sampling/importance_sampling_ratio/mean': 1.0153626203536987, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1479218304157257, 'clip_ratio/low_mean': 0.09406539052724838, 'clip_ratio/low_min': 0.09406539052724838, 'clip_ratio/high_mean': 0.08770995028316975, 'clip_ratio/high_max': 0.08770995028316975, 'clip_ratio/region_mean': 0.18177534081041813, 'reward_total_mean': 0.3942283093929291, 'reward_meter_mean': 0.8325039148330688, 'reward_meter_std': 0.2588021159172058, '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_lexical_plausibility_mean': 0.9751418232917786, 'reward_lexical_plausibility_std': 0.070309579372406, 'reward_repeat_penalty_mean': 0.9999389052391052, 'reward_repeat_penalty_std': 0.00017280207248404622, 'reward_near_duplicate_penalty_mean': 0.9999389052391052, 'reward_near_duplicate_penalty_std': 0.00017280207248404622, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4675000309944153, 'reward_judge_quality_std': 0.3998839259147644, 'reward_total_composite_mean': 0.3942283093929291, 'reward_total_composite_std': 0.35926198959350586, 'epoch': 0.02}

 14%|█▍        | 455/3300 [1:45:43<18:59:32, 24.03s/it]INFO 04-12 12:31:56 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 456/3300 [1:45:59<16:53:41, 21.39s/it]2026-04-12 12:32:11,167 | INFO | train_grpo_train | metrics_logged mode=train step=456

                                                       
{'loss': -0.1058, 'grad_norm': 5.302556991577148, 'learning_rate': 8.621212121212122e-06, 'num_tokens': 898002.0, 'completions/mean_length': 148.625, 'completions/min_length': 85.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 96.71428680419922, 'completions/min_terminated_length': 85.0, 'completions/max_terminated_length': 109.0, 'rewards/meter/mean': 0.611673891544342, 'rewards/meter/std': 0.3498254418373108, '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/lexical_plausibility/mean': 0.9578578472137451, 'rewards/lexical_plausibility/std': 0.10840871930122375, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.8715651035308838, 'rewards/repeat_penalty/std': 0.2136087566614151, 'rewards/near_duplicate_penalty/mean': 0.9509821534156799, 'rewards/near_duplicate_penalty/std': 0.06376034021377563, 'rewards/opening_diversity/mean': 0.9609375, 'rewards/opening_diversity/std': 0.08799287676811218, 'rewards/distinct_2/mean': 0.9321064949035645, 'rewards/distinct_2/std': 0.137295663356781, 'rewards/total_composite/mean': 0.2218300998210907, 'rewards/total_composite/std': 0.14927150309085846, 'reward': 0.2218300998210907, 'reward_std': 0.14927148818969727, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13501262664794922, 'sampling/sampling_logp_difference/max': 2.1519289016723633, 'sampling/importance_sampling_ratio/min': 0.11625968664884567, 'sampling/importance_sampling_ratio/mean': 1.021620273590088, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8340182527899742, 'clip_ratio/low_mean': 0.06189540959894657, 'clip_ratio/low_min': 0.06189540959894657, 'clip_ratio/high_mean': 0.05583109613507986, 'clip_ratio/high_max': 0.05583109613507986, 'clip_ratio/region_mean': 0.11772650573402643, 'reward_total_mean': 0.2218300998210907, 'reward_meter_mean': 0.611673891544342, 'reward_meter_std': 0.3498254418373108, '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_lexical_plausibility_mean': 0.9578578472137451, 'reward_lexical_plausibility_std': 0.10840871930122375, 'reward_repeat_penalty_mean': 0.8715651035308838, 'reward_repeat_penalty_std': 0.2136087566614151, 'reward_near_duplicate_penalty_mean': 0.9509821534156799, 'reward_near_duplicate_penalty_std': 0.06376034021377563, 'reward_opening_diversity_mean': 0.9609375, 'reward_opening_diversity_std': 0.08799287676811218, 'reward_distinct_2_mean': 0.9321064949035645, 'reward_distinct_2_std': 0.137295663356781, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.2218300998210907, 'reward_total_composite_std': 0.14927150309085846, 'epoch': 0.02}

 14%|█▍        | 456/3300 [1:45:59<16:53:41, 21.39s/it]INFO 04-12 12:32:11 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 457/3300 [1:46:11<14:41:36, 18.61s/it]2026-04-12 12:32:23,288 | INFO | train_grpo_train | metrics_logged mode=train step=457

                                                       
{'loss': 0.0642, 'grad_norm': 5.847695827484131, 'learning_rate': 8.618181818181819e-06, 'num_tokens': 901340.0, 'completions/mean_length': 188.25, 'completions/min_length': 169.0, 'completions/max_length': 240.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 188.25, 'completions/min_terminated_length': 169.0, 'completions/max_terminated_length': 240.0, 'rewards/meter/mean': 0.9877897500991821, 'rewards/meter/std': 0.006837547291070223, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.05892555043101311, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.17500001192092896, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/repeat_penalty/mean': 0.297836571931839, 'rewards/repeat_penalty/std': 0.29885193705558777, 'rewards/near_duplicate_penalty/mean': 0.8359514474868774, 'rewards/near_duplicate_penalty/std': 0.11799436062574387, 'rewards/opening_diversity/mean': 0.9732142686843872, 'rewards/opening_diversity/std': 0.04959750175476074, 'rewards/distinct_2/mean': 0.35675370693206787, 'rewards/distinct_2/std': 0.3196938633918762, 'rewards/total_composite/mean': 0.14082619547843933, 'rewards/total_composite/std': 0.05942106246948242, 'reward': 0.14082619547843933, 'reward_std': 0.05942106246948242, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09971191734075546, 'sampling/sampling_logp_difference/max': 1.7250299453735352, 'sampling/importance_sampling_ratio/min': 0.17816773056983948, 'sampling/importance_sampling_ratio/mean': 1.0115858316421509, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6942291334271431, 'clip_ratio/low_mean': 0.06989877298474312, 'clip_ratio/low_min': 0.06989877298474312, 'clip_ratio/high_mean': 0.02292899414896965, 'clip_ratio/high_max': 0.02292899414896965, 'clip_ratio/region_mean': 0.09282776713371277, 'reward_total_mean': 0.14082619547843933, 'reward_meter_mean': 0.9877897500991821, 'reward_meter_std': 0.006837547291070223, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.05892555043101311, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.297836571931839, 'reward_repeat_penalty_std': 0.29885193705558777, 'reward_near_duplicate_penalty_mean': 0.8359514474868774, 'reward_near_duplicate_penalty_std': 0.11799436062574387, 'reward_opening_diversity_mean': 0.9732142686843872, 'reward_opening_diversity_std': 0.04959750175476074, 'reward_distinct_2_mean': 0.35675370693206787, 'reward_distinct_2_std': 0.3196938633918762, 'reward_judge_quality_mean': 0.17500001192092896, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.14082619547843933, 'reward_total_composite_std': 0.05942106246948242, 'epoch': 0.02}

 14%|█▍        | 457/3300 [1:46:11<14:41:36, 18.61s/it]INFO 04-12 12:32:23 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 458/3300 [1:46:25<13:33:52, 17.18s/it]2026-04-12 12:32:37,148 | INFO | train_grpo_train | metrics_logged mode=train step=458

                                                       
{'loss': -0.0538, 'grad_norm': 3.4252231121063232, 'learning_rate': 8.615151515151516e-06, 'num_tokens': 902944.0, 'completions/mean_length': 162.5, 'completions/min_length': 44.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 46.0, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.2845132350921631, 'rewards/meter/std': 0.3603869676589966, '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/lexical_plausibility/mean': 0.9401000738143921, 'rewards/lexical_plausibility/std': 0.11121214181184769, 'rewards/judge_quality/mean': 0.5550000071525574, 'rewards/judge_quality/std': 0.3632394075393677, 'rewards/repeat_penalty/mean': 0.9974934458732605, 'rewards/repeat_penalty/std': 0.006550848484039307, 'rewards/near_duplicate_penalty/mean': 0.9974934458732605, 'rewards/near_duplicate_penalty/std': 0.006550848484039307, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.18996964395046234, 'rewards/total_composite/std': 0.23681482672691345, 'reward': 0.18996964395046234, 'reward_std': 0.23681481182575226, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17456616461277008, 'sampling/sampling_logp_difference/max': 2.0782382488250732, 'sampling/importance_sampling_ratio/min': 0.12515050172805786, 'sampling/importance_sampling_ratio/mean': 1.03304123878479, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.010149508714676, 'clip_ratio/low_mean': 0.0731565672904253, 'clip_ratio/low_min': 0.0731565672904253, 'clip_ratio/high_mean': 0.09311594255268574, 'clip_ratio/high_max': 0.09311594255268574, 'clip_ratio/region_mean': 0.16627250984311104, 'reward_total_mean': 0.18996964395046234, 'reward_meter_mean': 0.2845132350921631, 'reward_meter_std': 0.3603869676589966, '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_lexical_plausibility_mean': 0.9401000738143921, 'reward_lexical_plausibility_std': 0.11121214181184769, 'reward_repeat_penalty_mean': 0.9974934458732605, 'reward_repeat_penalty_std': 0.006550848484039307, 'reward_near_duplicate_penalty_mean': 0.9974934458732605, 'reward_near_duplicate_penalty_std': 0.006550848484039307, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5550000071525574, 'reward_judge_quality_std': 0.3632394075393677, 'reward_total_composite_mean': 0.18996964395046234, 'reward_total_composite_std': 0.23681482672691345, 'epoch': 0.02}

 14%|█▍        | 458/3300 [1:46:25<13:33:52, 17.18s/it]INFO 04-12 12:32:37 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 459/3300 [1:46:39<12:59:51, 16.47s/it]2026-04-12 12:32:51,962 | INFO | train_grpo_train | metrics_logged mode=train step=459

                                                       
{'loss': 0.0085, 'grad_norm': 1.7777029275894165, 'learning_rate': 8.612121212121213e-06, 'num_tokens': 904755.0, 'completions/mean_length': 306.375, 'completions/min_length': 46.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 100.75, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 232.0, 'rewards/meter/mean': 0.4911397695541382, 'rewards/meter/std': 0.42764201760292053, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.4432026445865631, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/lexical_plausibility/mean': 0.8601722121238708, 'rewards/lexical_plausibility/std': 0.12183038890361786, 'rewards/judge_quality/mean': 0.2587500214576721, 'rewards/judge_quality/std': 0.30898162722587585, 'rewards/repeat_penalty/mean': 0.8600711822509766, 'rewards/repeat_penalty/std': 0.34969615936279297, 'rewards/near_duplicate_penalty/mean': 0.9648240804672241, 'rewards/near_duplicate_penalty/std': 0.09662505984306335, 'rewards/opening_diversity/mean': 0.9838235378265381, 'rewards/opening_diversity/std': 0.03539901599287987, 'rewards/distinct_2/mean': 0.873298168182373, 'rewards/distinct_2/std': 0.35289791226387024, 'rewards/total_composite/mean': 0.17820200324058533, 'rewards/total_composite/std': 0.321745365858078, 'reward': 0.17820200324058533, 'reward_std': 0.3217453360557556, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11229118704795837, 'sampling/sampling_logp_difference/max': 1.5878190994262695, 'sampling/importance_sampling_ratio/min': 0.20437082648277283, 'sampling/importance_sampling_ratio/mean': 0.9947002530097961, 'sampling/importance_sampling_ratio/max': 1.7974423170089722, 'entropy': 0.5010434947907925, 'clip_ratio/low_mean': 0.028348806779831648, 'clip_ratio/low_min': 0.028348806779831648, 'clip_ratio/high_mean': 0.05631359852850437, 'clip_ratio/high_max': 0.05631359852850437, 'clip_ratio/region_mean': 0.08466240530833602, 'reward_total_mean': 0.17820200324058533, 'reward_meter_mean': 0.4911397695541382, 'reward_meter_std': 0.42764201760292053, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.4432026445865631, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_lexical_plausibility_mean': 0.8601722121238708, 'reward_lexical_plausibility_std': 0.12183038890361786, 'reward_repeat_penalty_mean': 0.8600711822509766, 'reward_repeat_penalty_std': 0.34969615936279297, 'reward_near_duplicate_penalty_mean': 0.9648240804672241, 'reward_near_duplicate_penalty_std': 0.09662505984306335, 'reward_opening_diversity_mean': 0.9838235378265381, 'reward_opening_diversity_std': 0.03539901599287987, 'reward_distinct_2_mean': 0.873298168182373, 'reward_distinct_2_std': 0.35289791226387024, 'reward_judge_quality_mean': 0.2587500214576721, 'reward_judge_quality_std': 0.30898162722587585, 'reward_total_composite_mean': 0.17820200324058533, 'reward_total_composite_std': 0.321745365858078, 'epoch': 0.02}

 14%|█▍        | 459/3300 [1:46:39<12:59:51, 16.47s/it]INFO 04-12 12:32:52 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 460/3300 [1:46:56<12:57:10, 16.42s/it]2026-04-12 12:33:08,261 | INFO | train_grpo_train | metrics_logged mode=train step=460

                                                       
{'loss': -0.0153, 'grad_norm': 5.404685974121094, 'learning_rate': 8.60909090909091e-06, 'num_tokens': 906512.0, 'completions/mean_length': 112.625, 'completions/min_length': 49.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 55.57143020629883, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.6130282878875732, 'rewards/meter/std': 0.406497985124588, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9545073509216309, 'rewards/lexical_plausibility/std': 0.1286727488040924, 'rewards/judge_quality/mean': 0.6474999785423279, 'rewards/judge_quality/std': 0.3930739760398865, 'rewards/repeat_penalty/mean': 0.9848916530609131, 'rewards/repeat_penalty/std': 0.01518380269408226, 'rewards/near_duplicate_penalty/mean': 0.9848916530609131, 'rewards/near_duplicate_penalty/std': 0.01518380269408226, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3965218663215637, 'rewards/total_composite/std': 0.4368768632411957, 'reward': 0.3965218663215637, 'reward_std': 0.43687689304351807, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16560642421245575, 'sampling/sampling_logp_difference/max': 1.5658135414123535, 'sampling/importance_sampling_ratio/min': 0.20891797542572021, 'sampling/importance_sampling_ratio/mean': 1.0113576650619507, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.271149955689907, 'clip_ratio/low_mean': 0.06828668527305126, 'clip_ratio/low_min': 0.06828668527305126, 'clip_ratio/high_mean': 0.05355507740750909, 'clip_ratio/high_max': 0.05355507740750909, 'clip_ratio/region_mean': 0.12184176268056035, 'reward_total_mean': 0.3965218663215637, 'reward_meter_mean': 0.6130282878875732, 'reward_meter_std': 0.406497985124588, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9545073509216309, 'reward_lexical_plausibility_std': 0.1286727488040924, 'reward_repeat_penalty_mean': 0.9848916530609131, 'reward_repeat_penalty_std': 0.01518380269408226, 'reward_near_duplicate_penalty_mean': 0.9848916530609131, 'reward_near_duplicate_penalty_std': 0.01518380269408226, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6474999785423279, 'reward_judge_quality_std': 0.3930739760398865, 'reward_total_composite_mean': 0.3965218663215637, 'reward_total_composite_std': 0.4368768632411957, 'epoch': 0.02}

 14%|█▍        | 460/3300 [1:46:56<12:57:10, 16.42s/it]INFO 04-12 12:33:08 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 461/3300 [1:47:05<11:17:26, 14.32s/it]2026-04-12 12:33:17,670 | INFO | train_grpo_train | metrics_logged mode=train step=461

                                                       
{'loss': 0.0929, 'grad_norm': 15.785818099975586, 'learning_rate': 8.606060606060606e-06, 'num_tokens': 908219.0, 'completions/mean_length': 46.375, 'completions/min_length': 40.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.375, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.6502194404602051, 'rewards/meter/std': 0.41814494132995605, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7637499570846558, 'rewards/judge_quality/std': 0.2332955002784729, 'rewards/repeat_penalty/mean': 0.9891205430030823, 'rewards/repeat_penalty/std': 0.024065691977739334, 'rewards/near_duplicate_penalty/mean': 0.9978618025779724, 'rewards/near_duplicate_penalty/std': 0.0032560399267822504, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9912587404251099, 'rewards/distinct_2/std': 0.024724015966057777, 'rewards/total_composite/mean': 0.49290597438812256, 'rewards/total_composite/std': 0.36434832215309143, 'reward': 0.49290597438812256, 'reward_std': 0.36434832215309143, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17635053396224976, 'sampling/sampling_logp_difference/max': 1.5859463214874268, 'sampling/importance_sampling_ratio/min': 0.20475393533706665, 'sampling/importance_sampling_ratio/mean': 1.0242173671722412, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4119069501757622, 'clip_ratio/low_mean': 0.07350852247327566, 'clip_ratio/low_min': 0.07350852247327566, 'clip_ratio/high_mean': 0.07539495546370745, 'clip_ratio/high_max': 0.07539495546370745, 'clip_ratio/region_mean': 0.1489034779369831, 'reward_total_mean': 0.49290597438812256, 'reward_meter_mean': 0.6502194404602051, 'reward_meter_std': 0.41814494132995605, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9891205430030823, 'reward_repeat_penalty_std': 0.024065691977739334, 'reward_near_duplicate_penalty_mean': 0.9978618025779724, 'reward_near_duplicate_penalty_std': 0.0032560399267822504, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9912587404251099, 'reward_distinct_2_std': 0.024724015966057777, 'reward_judge_quality_mean': 0.7637499570846558, 'reward_judge_quality_std': 0.2332955002784729, 'reward_total_composite_mean': 0.49290597438812256, 'reward_total_composite_std': 0.36434832215309143, 'epoch': 0.02}

 14%|█▍        | 461/3300 [1:47:05<11:17:26, 14.32s/it]INFO 04-12 12:33:17 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 462/3300 [1:47:19<11:09:09, 14.15s/it]2026-04-12 12:33:31,424 | INFO | train_grpo_train | metrics_logged mode=train step=462

                                                       
{'loss': -0.0128, 'grad_norm': 5.457734107971191, 'learning_rate': 8.603030303030303e-06, 'num_tokens': 910084.0, 'completions/mean_length': 139.125, 'completions/min_length': 63.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 85.85714721679688, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 126.0, 'rewards/meter/mean': 0.35627731680870056, 'rewards/meter/std': 0.3022831678390503, 'rewards/count_adherence/mean': 0.7916666865348816, 'rewards/count_adherence/std': 0.24800792336463928, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9872536659240723, 'rewards/lexical_plausibility/std': 0.036052118986845016, 'rewards/judge_quality/mean': 0.4087499976158142, 'rewards/judge_quality/std': 0.258370578289032, 'rewards/repeat_penalty/mean': 0.8636877536773682, 'rewards/repeat_penalty/std': 0.347133994102478, 'rewards/near_duplicate_penalty/mean': 0.929741621017456, 'rewards/near_duplicate_penalty/std': 0.16045944392681122, 'rewards/opening_diversity/mean': 0.9312499761581421, 'rewards/opening_diversity/std': 0.1944543719291687, 'rewards/distinct_2/mean': 0.8826597332954407, 'rewards/distinct_2/std': 0.33188846707344055, 'rewards/total_composite/mean': 0.10821406543254852, 'rewards/total_composite/std': 0.12190105766057968, 'reward': 0.10821406543254852, 'reward_std': 0.12190105020999908, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1310155987739563, 'sampling/sampling_logp_difference/max': 1.7909269332885742, 'sampling/importance_sampling_ratio/min': 0.1668054759502411, 'sampling/importance_sampling_ratio/mean': 1.0047672986984253, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.652204480022192, 'clip_ratio/low_mean': 0.06669470248743892, 'clip_ratio/low_min': 0.06669470248743892, 'clip_ratio/high_mean': 0.0298466170206666, 'clip_ratio/high_max': 0.0298466170206666, 'clip_ratio/region_mean': 0.09654131950810552, 'reward_total_mean': 0.10821406543254852, 'reward_meter_mean': 0.35627731680870056, 'reward_meter_std': 0.3022831678390503, 'reward_count_adherence_mean': 0.7916666865348816, 'reward_count_adherence_std': 0.24800792336463928, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9872536659240723, 'reward_lexical_plausibility_std': 0.036052118986845016, 'reward_repeat_penalty_mean': 0.8636877536773682, 'reward_repeat_penalty_std': 0.347133994102478, 'reward_near_duplicate_penalty_mean': 0.929741621017456, 'reward_near_duplicate_penalty_std': 0.16045944392681122, 'reward_opening_diversity_mean': 0.9312499761581421, 'reward_opening_diversity_std': 0.1944543719291687, 'reward_distinct_2_mean': 0.8826597332954407, 'reward_distinct_2_std': 0.33188846707344055, 'reward_judge_quality_mean': 0.4087499976158142, 'reward_judge_quality_std': 0.258370578289032, 'reward_total_composite_mean': 0.10821406543254852, 'reward_total_composite_std': 0.12190105766057968, 'epoch': 0.02}

 14%|█▍        | 462/3300 [1:47:19<11:09:09, 14.15s/it]INFO 04-12 12:33:31 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 463/3300 [1:47:34<11:16:45, 14.31s/it]2026-04-12 12:33:46,126 | INFO | train_grpo_train | metrics_logged mode=train step=463

                                                       
{'loss': -0.1033, 'grad_norm': 1.9380801916122437, 'learning_rate': 8.6e-06, 'num_tokens': 912070.0, 'completions/mean_length': 468.25, 'completions/min_length': 162.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.875, 'completions/mean_terminated_length': 162.0, 'completions/min_terminated_length': 162.0, 'completions/max_terminated_length': 162.0, 'rewards/meter/mean': 0.48026472330093384, 'rewards/meter/std': 0.3324054181575775, 'rewards/count_adherence/mean': 0.550000011920929, 'rewards/count_adherence/std': 0.3964124619960785, 'rewards/arabic_clean/mean': 0.125, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.7741320133209229, 'rewards/lexical_plausibility/std': 0.13840529322624207, 'rewards/judge_quality/mean': 0.125, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.9630327224731445, 'rewards/repeat_penalty/std': 0.08634717017412186, 'rewards/near_duplicate_penalty/mean': 0.9947658777236938, 'rewards/near_duplicate_penalty/std': 0.00745810242369771, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9995168447494507, 'rewards/distinct_2/std': 0.0013666537124663591, 'rewards/total_composite/mean': 0.021223342046141624, 'rewards/total_composite/std': 0.06002867966890335, 'reward': 0.021223342046141624, 'reward_std': 0.06002867594361305, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14795254170894623, 'sampling/sampling_logp_difference/max': 1.2214689254760742, 'sampling/importance_sampling_ratio/min': 0.29479679465293884, 'sampling/importance_sampling_ratio/mean': 1.0171161890029907, 'sampling/importance_sampling_ratio/max': 1.8904085159301758, 'entropy': 0.17437541484832764, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.014660493470728397, 'clip_ratio/high_max': 0.014660493470728397, 'clip_ratio/region_mean': 0.014660493470728397, 'reward_total_mean': 0.021223342046141624, 'reward_meter_mean': 0.48026472330093384, 'reward_meter_std': 0.3324054181575775, 'reward_count_adherence_mean': 0.550000011920929, 'reward_count_adherence_std': 0.3964124619960785, 'reward_arabic_clean_mean': 0.125, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.7741320133209229, 'reward_lexical_plausibility_std': 0.13840529322624207, 'reward_repeat_penalty_mean': 0.9630327224731445, 'reward_repeat_penalty_std': 0.08634717017412186, 'reward_near_duplicate_penalty_mean': 0.9947658777236938, 'reward_near_duplicate_penalty_std': 0.00745810242369771, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9995168447494507, 'reward_distinct_2_std': 0.0013666537124663591, 'reward_judge_quality_mean': 0.125, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.021223342046141624, 'reward_total_composite_std': 0.06002867966890335, 'epoch': 0.02}

 14%|█▍        | 463/3300 [1:47:34<11:16:45, 14.31s/it]INFO 04-12 12:33:46 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 464/3300 [1:47:47<11:06:56, 14.11s/it]2026-04-12 12:33:59,756 | INFO | train_grpo_train | metrics_logged mode=train step=464

                                                       
{'loss': -0.0916, 'grad_norm': 2.500302314758301, 'learning_rate': 8.596969696969698e-06, 'num_tokens': 913462.0, 'completions/mean_length': 224.0, 'completions/min_length': 45.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 51.20000076293945, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.815238893032074, 'rewards/meter/std': 0.3413713574409485, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.26726123690605164, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.8877477645874023, 'rewards/lexical_plausibility/std': 0.15713438391685486, 'rewards/judge_quality/mean': 0.41750001907348633, 'rewards/judge_quality/std': 0.3612182140350342, 'rewards/repeat_penalty/mean': 0.9581383466720581, 'rewards/repeat_penalty/std': 0.08907745778560638, 'rewards/near_duplicate_penalty/mean': 0.9975959062576294, 'rewards/near_duplicate_penalty/std': 0.006460277829319239, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9916387796401978, 'rewards/distinct_2/std': 0.0236490610986948, 'rewards/total_composite/mean': 0.32244977355003357, 'rewards/total_composite/std': 0.30975663661956787, 'reward': 0.32244977355003357, 'reward_std': 0.3097566068172455, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.167901411652565, 'sampling/sampling_logp_difference/max': 1.6881446838378906, 'sampling/importance_sampling_ratio/min': 0.1848621815443039, 'sampling/importance_sampling_ratio/mean': 1.042559027671814, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0568195208907127, 'clip_ratio/low_mean': 0.013888888992369175, 'clip_ratio/low_min': 0.013888888992369175, 'clip_ratio/high_mean': 0.08075993601232767, 'clip_ratio/high_max': 0.08075993601232767, 'clip_ratio/region_mean': 0.09464882500469685, 'reward_total_mean': 0.32244977355003357, 'reward_meter_mean': 0.815238893032074, 'reward_meter_std': 0.3413713574409485, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.26726123690605164, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.8877477645874023, 'reward_lexical_plausibility_std': 0.15713438391685486, 'reward_repeat_penalty_mean': 0.9581383466720581, 'reward_repeat_penalty_std': 0.08907745778560638, 'reward_near_duplicate_penalty_mean': 0.9975959062576294, 'reward_near_duplicate_penalty_std': 0.006460277829319239, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9916387796401978, 'reward_distinct_2_std': 0.0236490610986948, 'reward_judge_quality_mean': 0.41750001907348633, 'reward_judge_quality_std': 0.3612182140350342, 'reward_total_composite_mean': 0.32244977355003357, 'reward_total_composite_std': 0.30975663661956787, 'epoch': 0.02}

 14%|█▍        | 464/3300 [1:47:47<11:06:56, 14.11s/it]INFO 04-12 12:33:59 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 465/3300 [1:48:02<11:19:25, 14.38s/it]2026-04-12 12:34:14,768 | INFO | train_grpo_train | metrics_logged mode=train step=465

                                                       
{'loss': -0.0487, 'grad_norm': 3.8521788120269775, 'learning_rate': 8.593939393939395e-06, 'num_tokens': 914959.0, 'completions/mean_length': 160.125, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 42.833335876464844, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.5718278288841248, 'rewards/meter/std': 0.4065650701522827, '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/lexical_plausibility/mean': 0.9053156971931458, 'rewards/lexical_plausibility/std': 0.15994581580162048, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.18077214062213898, 'rewards/repeat_penalty/mean': 0.9709079265594482, 'rewards/repeat_penalty/std': 0.03222690522670746, 'rewards/near_duplicate_penalty/mean': 0.9794907569885254, 'rewards/near_duplicate_penalty/std': 0.022272756323218346, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9912587404251099, 'rewards/distinct_2/std': 0.024724015966057777, 'rewards/total_composite/mean': 0.1988687515258789, 'rewards/total_composite/std': 0.19370269775390625, 'reward': 0.1988687515258789, 'reward_std': 0.19370269775390625, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17526750266551971, 'sampling/sampling_logp_difference/max': 1.7957658767700195, 'sampling/importance_sampling_ratio/min': 0.16600026190280914, 'sampling/importance_sampling_ratio/mean': 1.0346271991729736, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.014502078294754, 'clip_ratio/low_mean': 0.03730158880352974, 'clip_ratio/low_min': 0.03730158880352974, 'clip_ratio/high_mean': 0.10084797907620668, 'clip_ratio/high_max': 0.10084797907620668, 'clip_ratio/region_mean': 0.13814956787973642, 'reward_total_mean': 0.1988687515258789, 'reward_meter_mean': 0.5718278288841248, 'reward_meter_std': 0.4065650701522827, '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_lexical_plausibility_mean': 0.9053156971931458, 'reward_lexical_plausibility_std': 0.15994581580162048, 'reward_repeat_penalty_mean': 0.9709079265594482, 'reward_repeat_penalty_std': 0.03222690522670746, 'reward_near_duplicate_penalty_mean': 0.9794907569885254, 'reward_near_duplicate_penalty_std': 0.022272756323218346, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9912587404251099, 'reward_distinct_2_std': 0.024724015966057777, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.18077214062213898, 'reward_total_composite_mean': 0.1988687515258789, 'reward_total_composite_std': 0.19370269775390625, 'epoch': 0.02}

 14%|█▍        | 465/3300 [1:48:02<11:19:25, 14.38s/it]INFO 04-12 12:34:15 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 466/3300 [1:48:16<11:12:19, 14.23s/it]2026-04-12 12:34:28,664 | INFO | train_grpo_train | metrics_logged mode=train step=466

                                                       
{'loss': 0.0068, 'grad_norm': 6.118508815765381, 'learning_rate': 8.590909090909092e-06, 'num_tokens': 917132.0, 'completions/mean_length': 164.625, 'completions/min_length': 95.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 115.00000762939453, 'completions/min_terminated_length': 95.0, 'completions/max_terminated_length': 130.0, 'rewards/meter/mean': 0.6192550659179688, 'rewards/meter/std': 0.3906235992908478, 'rewards/count_adherence/mean': 0.78125, 'rewards/count_adherence/std': 0.24775780737400055, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9633049964904785, 'rewards/lexical_plausibility/std': 0.10378912091255188, 'rewards/judge_quality/mean': 0.3412500023841858, 'rewards/judge_quality/std': 0.2564280927181244, 'rewards/repeat_penalty/mean': 0.7339690923690796, 'rewards/repeat_penalty/std': 0.2971440255641937, 'rewards/near_duplicate_penalty/mean': 0.936042308807373, 'rewards/near_duplicate_penalty/std': 0.06508824229240417, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.8219245076179504, 'rewards/distinct_2/std': 0.21221967041492462, 'rewards/total_composite/mean': 0.22211773693561554, 'rewards/total_composite/std': 0.2933432161808014, 'reward': 0.22211773693561554, 'reward_std': 0.2933432161808014, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15400467813014984, 'sampling/sampling_logp_difference/max': 2.870518207550049, 'sampling/importance_sampling_ratio/min': 0.05666955187916756, 'sampling/importance_sampling_ratio/mean': 1.0181410312652588, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9257434085011482, 'clip_ratio/low_mean': 0.06992149259895086, 'clip_ratio/low_min': 0.06992149259895086, 'clip_ratio/high_mean': 0.06821741908788681, 'clip_ratio/high_max': 0.06821741908788681, 'clip_ratio/region_mean': 0.13813891168683767, 'reward_total_mean': 0.22211773693561554, 'reward_meter_mean': 0.6192550659179688, 'reward_meter_std': 0.3906235992908478, 'reward_count_adherence_mean': 0.78125, 'reward_count_adherence_std': 0.24775780737400055, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9633049964904785, 'reward_lexical_plausibility_std': 0.10378912091255188, 'reward_repeat_penalty_mean': 0.7339690923690796, 'reward_repeat_penalty_std': 0.2971440255641937, 'reward_near_duplicate_penalty_mean': 0.936042308807373, 'reward_near_duplicate_penalty_std': 0.06508824229240417, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.8219245076179504, 'reward_distinct_2_std': 0.21221967041492462, 'reward_judge_quality_mean': 0.3412500023841858, 'reward_judge_quality_std': 0.2564280927181244, 'reward_total_composite_mean': 0.22211773693561554, 'reward_total_composite_std': 0.2933432161808014, 'epoch': 0.02}

 14%|█▍        | 466/3300 [1:48:16<11:12:19, 14.23s/it]INFO 04-12 12:34:28 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 467/3300 [1:48:28<10:45:04, 13.66s/it]2026-04-12 12:34:40,994 | INFO | train_grpo_train | metrics_logged mode=train step=467

                                                       
{'loss': 0.0374, 'grad_norm': 3.205953598022461, 'learning_rate': 8.587878787878788e-06, 'num_tokens': 921223.0, 'completions/mean_length': 301.375, 'completions/min_length': 264.0, 'completions/max_length': 386.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 301.375, 'completions/min_terminated_length': 264.0, 'completions/max_terminated_length': 386.0, 'rewards/meter/mean': 0.9814234972000122, 'rewards/meter/std': 0.013453001156449318, 'rewards/count_adherence/mean': 0.7222222089767456, 'rewards/count_adherence/std': 0.11878278106451035, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.13750000298023224, 'rewards/judge_quality/std': 0.0353553406894207, 'rewards/repeat_penalty/mean': 0.005827170331031084, 'rewards/repeat_penalty/std': 0.016481727361679077, 'rewards/near_duplicate_penalty/mean': 0.5068979263305664, 'rewards/near_duplicate_penalty/std': 0.1939096450805664, 'rewards/opening_diversity/mean': 0.589586615562439, 'rewards/opening_diversity/std': 0.2773093581199646, 'rewards/distinct_2/mean': 0.017191141843795776, 'rewards/distinct_2/std': 0.04862389340996742, 'rewards/total_composite/mean': 0.09664815664291382, 'rewards/total_composite/std': 0.02899831160902977, 'reward': 0.09664815664291382, 'reward_std': 0.02899831160902977, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0416940413415432, 'sampling/sampling_logp_difference/max': 2.6657280921936035, 'sampling/importance_sampling_ratio/min': 0.06954869627952576, 'sampling/importance_sampling_ratio/mean': 1.0035876035690308, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3000090792775154, 'clip_ratio/low_mean': 0.00976654514670372, 'clip_ratio/low_min': 0.00976654514670372, 'clip_ratio/high_mean': 0.02837038761936128, 'clip_ratio/high_max': 0.02837038761936128, 'clip_ratio/region_mean': 0.038136932766065, 'reward_total_mean': 0.09664815664291382, 'reward_meter_mean': 0.9814234972000122, 'reward_meter_std': 0.013453001156449318, 'reward_count_adherence_mean': 0.7222222089767456, 'reward_count_adherence_std': 0.11878278106451035, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.005827170331031084, 'reward_repeat_penalty_std': 0.016481727361679077, 'reward_near_duplicate_penalty_mean': 0.5068979263305664, 'reward_near_duplicate_penalty_std': 0.1939096450805664, 'reward_opening_diversity_mean': 0.589586615562439, 'reward_opening_diversity_std': 0.2773093581199646, 'reward_distinct_2_mean': 0.017191141843795776, 'reward_distinct_2_std': 0.04862389340996742, 'reward_judge_quality_mean': 0.13750000298023224, 'reward_judge_quality_std': 0.0353553406894207, 'reward_total_composite_mean': 0.09664815664291382, 'reward_total_composite_std': 0.02899831160902977, 'epoch': 0.02}

 14%|█▍        | 467/3300 [1:48:28<10:45:04, 13.66s/it]INFO 04-12 12:34:41 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 468/3300 [1:48:38<9:48:20, 12.46s/it] 2026-04-12 12:34:50,660 | INFO | train_grpo_train | metrics_logged mode=train step=468

                                                      
{'loss': 0.132, 'grad_norm': 14.356237411499023, 'learning_rate': 8.584848484848485e-06, 'num_tokens': 923162.0, 'completions/mean_length': 84.375, 'completions/min_length': 58.0, 'completions/max_length': 105.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 84.375, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.7646868824958801, 'rewards/meter/std': 0.26857540011405945, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.2357022613286972, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9935064911842346, 'rewards/lexical_plausibility/std': 0.012782488018274307, 'rewards/judge_quality/mean': 0.47499996423721313, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/repeat_penalty/mean': 0.9549342393875122, 'rewards/repeat_penalty/std': 0.03803550451993942, 'rewards/near_duplicate_penalty/mean': 0.9848763346672058, 'rewards/near_duplicate_penalty/std': 0.007629488594830036, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9694910645484924, 'rewards/distinct_2/std': 0.03459250554442406, 'rewards/total_composite/mean': 0.2837463915348053, 'rewards/total_composite/std': 0.13871093094348907, 'reward': 0.2837463915348053, 'reward_std': 0.13871093094348907, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15805473923683167, 'sampling/sampling_logp_difference/max': 1.5223498344421387, 'sampling/importance_sampling_ratio/min': 0.21819856762886047, 'sampling/importance_sampling_ratio/mean': 1.0121657848358154, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0732149854302406, 'clip_ratio/low_mean': 0.03880208358168602, 'clip_ratio/low_min': 0.03880208358168602, 'clip_ratio/high_mean': 0.12570246774703264, 'clip_ratio/high_max': 0.12570246774703264, 'clip_ratio/region_mean': 0.16450455132871866, 'reward_total_mean': 0.2837463915348053, 'reward_meter_mean': 0.7646868824958801, 'reward_meter_std': 0.26857540011405945, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.2357022613286972, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9935064911842346, 'reward_lexical_plausibility_std': 0.012782488018274307, 'reward_repeat_penalty_mean': 0.9549342393875122, 'reward_repeat_penalty_std': 0.03803550451993942, 'reward_near_duplicate_penalty_mean': 0.9848763346672058, 'reward_near_duplicate_penalty_std': 0.007629488594830036, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9694910645484924, 'reward_distinct_2_std': 0.03459250554442406, 'reward_judge_quality_mean': 0.47499996423721313, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.2837463915348053, 'reward_total_composite_std': 0.13871093094348907, 'epoch': 0.02}

 14%|█▍        | 468/3300 [1:48:38<9:48:20, 12.46s/it]INFO 04-12 12:34:50 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 469/3300 [1:48:52<10:13:30, 13.00s/it]2026-04-12 12:35:04,913 | INFO | train_grpo_train | metrics_logged mode=train step=469

                                                       
{'loss': -0.1855, 'grad_norm': 2.868105411529541, 'learning_rate': 8.581818181818183e-06, 'num_tokens': 925255.0, 'completions/mean_length': 281.625, 'completions/min_length': 122.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 143.40000915527344, 'completions/min_terminated_length': 122.0, 'completions/max_terminated_length': 176.0, 'rewards/meter/mean': 0.787610650062561, 'rewards/meter/std': 0.3252897262573242, 'rewards/count_adherence/mean': 0.78125, 'rewards/count_adherence/std': 0.1602174937725067, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.8840222358703613, 'rewards/lexical_plausibility/std': 0.16291122138500214, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.2199837565422058, 'rewards/repeat_penalty/mean': 0.8537392020225525, 'rewards/repeat_penalty/std': 0.21768257021903992, 'rewards/near_duplicate_penalty/mean': 0.9621591567993164, 'rewards/near_duplicate_penalty/std': 0.04488173872232437, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.906207799911499, 'rewards/distinct_2/std': 0.15854863822460175, 'rewards/total_composite/mean': 0.12211199104785919, 'rewards/total_composite/std': 0.13145306706428528, 'reward': 0.12211199104785919, 'reward_std': 0.13145306706428528, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12535236775875092, 'sampling/sampling_logp_difference/max': 1.8409175872802734, 'sampling/importance_sampling_ratio/min': 0.1586717665195465, 'sampling/importance_sampling_ratio/mean': 1.0280526876449585, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5625474229454994, 'clip_ratio/low_mean': 0.012295082211494446, 'clip_ratio/low_min': 0.012295082211494446, 'clip_ratio/high_mean': 0.06628303416073322, 'clip_ratio/high_max': 0.06628303416073322, 'clip_ratio/region_mean': 0.07857811637222767, 'reward_total_mean': 0.12211199104785919, 'reward_meter_mean': 0.787610650062561, 'reward_meter_std': 0.3252897262573242, 'reward_count_adherence_mean': 0.78125, 'reward_count_adherence_std': 0.1602174937725067, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.8840222358703613, 'reward_lexical_plausibility_std': 0.16291122138500214, 'reward_repeat_penalty_mean': 0.8537392020225525, 'reward_repeat_penalty_std': 0.21768257021903992, 'reward_near_duplicate_penalty_mean': 0.9621591567993164, 'reward_near_duplicate_penalty_std': 0.04488173872232437, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.906207799911499, 'reward_distinct_2_std': 0.15854863822460175, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.2199837565422058, 'reward_total_composite_mean': 0.12211199104785919, 'reward_total_composite_std': 0.13145306706428528, 'epoch': 0.02}

 14%|█▍        | 469/3300 [1:48:52<10:13:30, 13.00s/it]INFO 04-12 12:35:05 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 470/3300 [1:49:07<10:35:16, 13.47s/it]2026-04-12 12:35:19,473 | INFO | train_grpo_train | metrics_logged mode=train step=470

                                                       
{'loss': -0.1159, 'grad_norm': 3.014456033706665, 'learning_rate': 8.57878787878788e-06, 'num_tokens': 928597.0, 'completions/mean_length': 271.75, 'completions/min_length': 210.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 237.4285888671875, 'completions/min_terminated_length': 210.0, 'completions/max_terminated_length': 251.0, 'rewards/meter/mean': 0.941405177116394, 'rewards/meter/std': 0.09066831320524216, 'rewards/count_adherence/mean': 0.8035714030265808, 'rewards/count_adherence/std': 0.10628911107778549, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9671759605407715, 'rewards/lexical_plausibility/std': 0.09284044802188873, 'rewards/judge_quality/mean': 0.1875, 'rewards/judge_quality/std': 0.1060660183429718, 'rewards/repeat_penalty/mean': 0.12993720173835754, 'rewards/repeat_penalty/std': 0.18206985294818878, 'rewards/near_duplicate_penalty/mean': 0.7461811900138855, 'rewards/near_duplicate_penalty/std': 0.12352963536977768, 'rewards/opening_diversity/mean': 0.7981770634651184, 'rewards/opening_diversity/std': 0.2740727663040161, 'rewards/distinct_2/mean': 0.255423367023468, 'rewards/distinct_2/std': 0.277616024017334, 'rewards/total_composite/mean': 0.14619973301887512, 'rewards/total_composite/std': 0.10242842882871628, 'reward': 0.14619973301887512, 'reward_std': 0.10242842137813568, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07308058440685272, 'sampling/sampling_logp_difference/max': 1.3938860893249512, 'sampling/importance_sampling_ratio/min': 0.24810925126075745, 'sampling/importance_sampling_ratio/mean': 1.0120410919189453, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5316671617329121, 'clip_ratio/low_mean': 0.03577795741148293, 'clip_ratio/low_min': 0.03577795741148293, 'clip_ratio/high_mean': 0.021479230374097824, 'clip_ratio/high_max': 0.021479230374097824, 'clip_ratio/region_mean': 0.057257187785580754, 'reward_total_mean': 0.14619973301887512, 'reward_meter_mean': 0.941405177116394, 'reward_meter_std': 0.09066831320524216, 'reward_count_adherence_mean': 0.8035714030265808, 'reward_count_adherence_std': 0.10628911107778549, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9671759605407715, 'reward_lexical_plausibility_std': 0.09284044802188873, 'reward_repeat_penalty_mean': 0.12993720173835754, 'reward_repeat_penalty_std': 0.18206985294818878, 'reward_near_duplicate_penalty_mean': 0.7461811900138855, 'reward_near_duplicate_penalty_std': 0.12352963536977768, 'reward_opening_diversity_mean': 0.7981770634651184, 'reward_opening_diversity_std': 0.2740727663040161, 'reward_distinct_2_mean': 0.255423367023468, 'reward_distinct_2_std': 0.277616024017334, 'reward_judge_quality_mean': 0.1875, 'reward_judge_quality_std': 0.1060660183429718, 'reward_total_composite_mean': 0.14619973301887512, 'reward_total_composite_std': 0.10242842882871628, 'epoch': 0.02}

 14%|█▍        | 470/3300 [1:49:07<10:35:16, 13.47s/it]INFO 04-12 12:35:19 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 471/3300 [1:49:20<10:29:29, 13.35s/it]2026-04-12 12:35:32,547 | INFO | train_grpo_train | metrics_logged mode=train step=471

                                                       
{'loss': -0.02, 'grad_norm': 3.8852856159210205, 'learning_rate': 8.575757575757575e-06, 'num_tokens': 930057.0, 'completions/mean_length': 148.5, 'completions/min_length': 19.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 27.33333396911621, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.5910894274711609, 'rewards/meter/std': 0.4692842960357666, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.4629100561141968, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.8728610277175903, 'rewards/lexical_plausibility/std': 0.19085237383842468, 'rewards/judge_quality/mean': 0.5012500286102295, 'rewards/judge_quality/std': 0.3744877576828003, 'rewards/repeat_penalty/mean': 0.8007950782775879, 'rewards/repeat_penalty/std': 0.11195925623178482, 'rewards/near_duplicate_penalty/mean': 0.9861470460891724, 'rewards/near_duplicate_penalty/std': 0.026621105149388313, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3510221242904663, 'rewards/total_composite/std': 0.37824493646621704, 'reward': 0.3510221242904663, 'reward_std': 0.37824493646621704, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1798359453678131, 'sampling/sampling_logp_difference/max': 1.9128549098968506, 'sampling/importance_sampling_ratio/min': 0.14765822887420654, 'sampling/importance_sampling_ratio/mean': 1.033026099205017, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0123971924185753, 'clip_ratio/low_mean': 0.08714237529784441, 'clip_ratio/low_min': 0.08714237529784441, 'clip_ratio/high_mean': 0.0580357164144516, 'clip_ratio/high_max': 0.0580357164144516, 'clip_ratio/region_mean': 0.145178091712296, 'reward_total_mean': 0.3510221242904663, 'reward_meter_mean': 0.5910894274711609, 'reward_meter_std': 0.4692842960357666, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.4629100561141968, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.8728610277175903, 'reward_lexical_plausibility_std': 0.19085237383842468, 'reward_repeat_penalty_mean': 0.8007950782775879, 'reward_repeat_penalty_std': 0.11195925623178482, 'reward_near_duplicate_penalty_mean': 0.9861470460891724, 'reward_near_duplicate_penalty_std': 0.026621105149388313, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5012500286102295, 'reward_judge_quality_std': 0.3744877576828003, 'reward_total_composite_mean': 0.3510221242904663, 'reward_total_composite_std': 0.37824493646621704, 'epoch': 0.02}

 14%|█▍        | 471/3300 [1:49:20<10:29:29, 13.35s/it]INFO 04-12 12:35:32 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 472/3300 [1:49:29<9:27:04, 12.03s/it] 2026-04-12 12:35:41,512 | INFO | train_grpo_train | metrics_logged mode=train step=472

                                                      
{'loss': 0.0709, 'grad_norm': 17.05718231201172, 'learning_rate': 8.572727272727274e-06, 'num_tokens': 931810.0, 'completions/mean_length': 58.125, 'completions/min_length': 48.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.125, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.48700374364852905, 'rewards/meter/std': 0.36860910058021545, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9787582159042358, 'rewards/lexical_plausibility/std': 0.04981458932161331, 'rewards/judge_quality/mean': 0.5087499618530273, 'rewards/judge_quality/std': 0.23092593252658844, 'rewards/repeat_penalty/mean': 0.9842442274093628, 'rewards/repeat_penalty/std': 0.020871134474873543, 'rewards/near_duplicate_penalty/mean': 0.9842442274093628, 'rewards/near_duplicate_penalty/std': 0.020871134474873543, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.23971016705036163, 'rewards/total_composite/std': 0.19003631174564362, 'reward': 0.23971016705036163, 'reward_std': 0.19003631174564362, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18038438260555267, 'sampling/sampling_logp_difference/max': 2.1869399547576904, 'sampling/importance_sampling_ratio/min': 0.11225974559783936, 'sampling/importance_sampling_ratio/mean': 1.0043736696243286, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9426776245236397, 'clip_ratio/low_mean': 0.07151182740926743, 'clip_ratio/low_min': 0.07151182740926743, 'clip_ratio/high_mean': 0.09448559768497944, 'clip_ratio/high_max': 0.09448559768497944, 'clip_ratio/region_mean': 0.16599742509424686, 'reward_total_mean': 0.23971016705036163, 'reward_meter_mean': 0.48700374364852905, 'reward_meter_std': 0.36860910058021545, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9787582159042358, 'reward_lexical_plausibility_std': 0.04981458932161331, 'reward_repeat_penalty_mean': 0.9842442274093628, 'reward_repeat_penalty_std': 0.020871134474873543, 'reward_near_duplicate_penalty_mean': 0.9842442274093628, 'reward_near_duplicate_penalty_std': 0.020871134474873543, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5087499618530273, 'reward_judge_quality_std': 0.23092593252658844, 'reward_total_composite_mean': 0.23971016705036163, 'reward_total_composite_std': 0.19003631174564362, 'epoch': 0.02}

 14%|█▍        | 472/3300 [1:49:29<9:27:04, 12.03s/it]INFO 04-12 12:35:41 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 473/3300 [1:49:37<8:36:23, 10.96s/it]2026-04-12 12:35:49,961 | INFO | train_grpo_train | metrics_logged mode=train step=473

                                                      
{'loss': 0.0744, 'grad_norm': 22.04007339477539, 'learning_rate': 8.56969696969697e-06, 'num_tokens': 933337.0, 'completions/mean_length': 22.875, 'completions/min_length': 20.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.875, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.9542794227600098, 'rewards/meter/std': 0.08132888376712799, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9632352590560913, 'rewards/lexical_plausibility/std': 0.10398628562688828, 'rewards/judge_quality/mean': 0.5299999713897705, 'rewards/judge_quality/std': 0.24512389302253723, 'rewards/repeat_penalty/mean': 0.717903733253479, 'rewards/repeat_penalty/std': 0.09078184515237808, 'rewards/near_duplicate_penalty/mean': 0.957205057144165, 'rewards/near_duplicate_penalty/std': 0.12104246020317078, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5053606033325195, 'rewards/total_composite/std': 0.2389005422592163, 'reward': 0.5053606033325195, 'reward_std': 0.2389005422592163, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19822415709495544, 'sampling/sampling_logp_difference/max': 2.484585762023926, 'sampling/importance_sampling_ratio/min': 0.08336007595062256, 'sampling/importance_sampling_ratio/mean': 1.0645005702972412, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1023565083742142, 'clip_ratio/low_mean': 0.14330357359722257, 'clip_ratio/low_min': 0.14330357359722257, 'clip_ratio/high_mean': 0.0535714291036129, 'clip_ratio/high_max': 0.0535714291036129, 'clip_ratio/region_mean': 0.19687500270083547, 'reward_total_mean': 0.5053606033325195, 'reward_meter_mean': 0.9542794227600098, 'reward_meter_std': 0.08132888376712799, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9632352590560913, 'reward_lexical_plausibility_std': 0.10398628562688828, 'reward_repeat_penalty_mean': 0.717903733253479, 'reward_repeat_penalty_std': 0.09078184515237808, 'reward_near_duplicate_penalty_mean': 0.957205057144165, 'reward_near_duplicate_penalty_std': 0.12104246020317078, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5299999713897705, 'reward_judge_quality_std': 0.24512389302253723, 'reward_total_composite_mean': 0.5053606033325195, 'reward_total_composite_std': 0.2389005422592163, 'epoch': 0.02}

 14%|█▍        | 473/3300 [1:49:37<8:36:23, 10.96s/it]INFO 04-12 12:35:50 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 474/3300 [1:49:52<9:26:57, 12.04s/it]2026-04-12 12:36:04,513 | INFO | train_grpo_train | metrics_logged mode=train step=474

                                                      
{'loss': -0.3999, 'grad_norm': 2.3125405311584473, 'learning_rate': 8.566666666666667e-06, 'num_tokens': 937105.0, 'completions/mean_length': 442.0, 'completions/min_length': 333.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 400.0, 'completions/min_terminated_length': 333.0, 'completions/max_terminated_length': 424.0, 'rewards/meter/mean': 0.7116121053695679, 'rewards/meter/std': 0.4438537359237671, 'rewards/count_adherence/mean': 0.4318181872367859, 'rewards/count_adherence/std': 0.3392818570137024, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.915495753288269, 'rewards/lexical_plausibility/std': 0.12496782839298248, 'rewards/judge_quality/mean': 0.125, 'rewards/judge_quality/std': 0.0707106813788414, 'rewards/repeat_penalty/mean': 0.37211185693740845, 'rewards/repeat_penalty/std': 0.46044665575027466, 'rewards/near_duplicate_penalty/mean': 0.8078860640525818, 'rewards/near_duplicate_penalty/std': 0.21712858974933624, 'rewards/opening_diversity/mean': 0.8451547026634216, 'rewards/opening_diversity/std': 0.2363804429769516, 'rewards/distinct_2/mean': 0.42280104756355286, 'rewards/distinct_2/std': 0.48468995094299316, 'rewards/total_composite/mean': 0.06848539412021637, 'rewards/total_composite/std': 0.05783623829483986, 'reward': 0.06848539412021637, 'reward_std': 0.05783623456954956, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07168090343475342, 'sampling/sampling_logp_difference/max': 1.3686151504516602, 'sampling/importance_sampling_ratio/min': 0.25445911288261414, 'sampling/importance_sampling_ratio/mean': 1.0107157230377197, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4093349613249302, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.039252737537026405, 'clip_ratio/high_max': 0.039252737537026405, 'clip_ratio/region_mean': 0.039252737537026405, 'reward_total_mean': 0.06848539412021637, 'reward_meter_mean': 0.7116121053695679, 'reward_meter_std': 0.4438537359237671, 'reward_count_adherence_mean': 0.4318181872367859, 'reward_count_adherence_std': 0.3392818570137024, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.915495753288269, 'reward_lexical_plausibility_std': 0.12496782839298248, 'reward_repeat_penalty_mean': 0.37211185693740845, 'reward_repeat_penalty_std': 0.46044665575027466, 'reward_near_duplicate_penalty_mean': 0.8078860640525818, 'reward_near_duplicate_penalty_std': 0.21712858974933624, 'reward_opening_diversity_mean': 0.8451547026634216, 'reward_opening_diversity_std': 0.2363804429769516, 'reward_distinct_2_mean': 0.42280104756355286, 'reward_distinct_2_std': 0.48468995094299316, 'reward_judge_quality_mean': 0.125, 'reward_judge_quality_std': 0.0707106813788414, 'reward_total_composite_mean': 0.06848539412021637, 'reward_total_composite_std': 0.05783623829483986, 'epoch': 0.02}

 14%|█▍        | 474/3300 [1:49:52<9:26:57, 12.04s/it]INFO 04-12 12:36:04 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 475/3300 [1:50:06<9:59:18, 12.73s/it]2026-04-12 12:36:18,860 | INFO | train_grpo_train | metrics_logged mode=train step=475

                                                      
{'loss': -0.1654, 'grad_norm': 2.6581056118011475, 'learning_rate': 8.563636363636364e-06, 'num_tokens': 939308.0, 'completions/mean_length': 272.375, 'completions/min_length': 114.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 128.60000610351562, 'completions/min_terminated_length': 114.0, 'completions/max_terminated_length': 152.0, 'rewards/meter/mean': 0.6184227466583252, 'rewards/meter/std': 0.4197257161140442, 'rewards/count_adherence/mean': 0.71875, 'rewards/count_adherence/std': 0.38816189765930176, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.8906276226043701, 'rewards/lexical_plausibility/std': 0.15246494114398956, 'rewards/judge_quality/mean': 0.29374998807907104, 'rewards/judge_quality/std': 0.19537052512168884, 'rewards/repeat_penalty/mean': 0.9596143364906311, 'rewards/repeat_penalty/std': 0.08553502708673477, 'rewards/near_duplicate_penalty/mean': 0.9945335388183594, 'rewards/near_duplicate_penalty/std': 0.00535243097692728, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9963017702102661, 'rewards/distinct_2/std': 0.01046015229076147, 'rewards/total_composite/mean': 0.19763284921646118, 'rewards/total_composite/std': 0.18747515976428986, 'reward': 0.19763284921646118, 'reward_std': 0.18747515976428986, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17012016475200653, 'sampling/sampling_logp_difference/max': 3.704521894454956, 'sampling/importance_sampling_ratio/min': 0.024611981585621834, 'sampling/importance_sampling_ratio/mean': 0.9970017075538635, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7514044344425201, 'clip_ratio/low_mean': 0.014860140159726143, 'clip_ratio/low_min': 0.014860140159726143, 'clip_ratio/high_mean': 0.08834486082196236, 'clip_ratio/high_max': 0.08834486082196236, 'clip_ratio/region_mean': 0.1032050009816885, 'reward_total_mean': 0.19763284921646118, 'reward_meter_mean': 0.6184227466583252, 'reward_meter_std': 0.4197257161140442, 'reward_count_adherence_mean': 0.71875, 'reward_count_adherence_std': 0.38816189765930176, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.8906276226043701, 'reward_lexical_plausibility_std': 0.15246494114398956, 'reward_repeat_penalty_mean': 0.9596143364906311, 'reward_repeat_penalty_std': 0.08553502708673477, 'reward_near_duplicate_penalty_mean': 0.9945335388183594, 'reward_near_duplicate_penalty_std': 0.00535243097692728, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9963017702102661, 'reward_distinct_2_std': 0.01046015229076147, 'reward_judge_quality_mean': 0.29374998807907104, 'reward_judge_quality_std': 0.19537052512168884, 'reward_total_composite_mean': 0.19763284921646118, 'reward_total_composite_std': 0.18747515976428986, 'epoch': 0.02}

 14%|█▍        | 475/3300 [1:50:06<9:59:18, 12.73s/it]INFO 04-12 12:36:19 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 476/3300 [1:50:21<10:22:03, 13.22s/it]2026-04-12 12:36:33,212 | INFO | train_grpo_train | metrics_logged mode=train step=476

                                                       
{'loss': -0.1407, 'grad_norm': 4.104028224945068, 'learning_rate': 8.560606060606062e-06, 'num_tokens': 941664.0, 'completions/mean_length': 253.5, 'completions/min_length': 145.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 167.33334350585938, 'completions/min_terminated_length': 145.0, 'completions/max_terminated_length': 188.0, 'rewards/meter/mean': 0.6942492127418518, 'rewards/meter/std': 0.43733933568000793, 'rewards/count_adherence/mean': 0.7708333134651184, 'rewards/count_adherence/std': 0.32043495774269104, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9109879732131958, 'rewards/lexical_plausibility/std': 0.16487261652946472, 'rewards/judge_quality/mean': 0.22500000894069672, 'rewards/judge_quality/std': 0.12817399203777313, 'rewards/repeat_penalty/mean': 0.4662206172943115, 'rewards/repeat_penalty/std': 0.36342477798461914, 'rewards/near_duplicate_penalty/mean': 0.8423084020614624, 'rewards/near_duplicate_penalty/std': 0.12640371918678284, 'rewards/opening_diversity/mean': 0.8956043720245361, 'rewards/opening_diversity/std': 0.1995983123779297, 'rewards/distinct_2/mean': 0.5765867233276367, 'rewards/distinct_2/std': 0.33712494373321533, 'rewards/total_composite/mean': 0.13432076573371887, 'rewards/total_composite/std': 0.13123878836631775, 'reward': 0.13432076573371887, 'reward_std': 0.13123878836631775, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1319139301776886, 'sampling/sampling_logp_difference/max': 2.1389307975769043, 'sampling/importance_sampling_ratio/min': 0.174083411693573, 'sampling/importance_sampling_ratio/mean': 1.0293654203414917, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7540062218904495, 'clip_ratio/low_mean': 0.03421391034498811, 'clip_ratio/low_min': 0.03421391034498811, 'clip_ratio/high_mean': 0.04104610625654459, 'clip_ratio/high_max': 0.04104610625654459, 'clip_ratio/region_mean': 0.0752600166015327, 'reward_total_mean': 0.13432076573371887, 'reward_meter_mean': 0.6942492127418518, 'reward_meter_std': 0.43733933568000793, 'reward_count_adherence_mean': 0.7708333134651184, 'reward_count_adherence_std': 0.32043495774269104, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9109879732131958, 'reward_lexical_plausibility_std': 0.16487261652946472, 'reward_repeat_penalty_mean': 0.4662206172943115, 'reward_repeat_penalty_std': 0.36342477798461914, 'reward_near_duplicate_penalty_mean': 0.8423084020614624, 'reward_near_duplicate_penalty_std': 0.12640371918678284, 'reward_opening_diversity_mean': 0.8956043720245361, 'reward_opening_diversity_std': 0.1995983123779297, 'reward_distinct_2_mean': 0.5765867233276367, 'reward_distinct_2_std': 0.33712494373321533, 'reward_judge_quality_mean': 0.22500000894069672, 'reward_judge_quality_std': 0.12817399203777313, 'reward_total_composite_mean': 0.13432076573371887, 'reward_total_composite_std': 0.13123878836631775, 'epoch': 0.02}

 14%|█▍        | 476/3300 [1:50:21<10:22:03, 13.22s/it]INFO 04-12 12:36:33 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 477/3300 [1:50:34<10:20:30, 13.19s/it]2026-04-12 12:36:46,332 | INFO | train_grpo_train | metrics_logged mode=train step=477

                                                       
{'loss': -0.0733, 'grad_norm': 5.797103404998779, 'learning_rate': 8.557575757575757e-06, 'num_tokens': 943334.0, 'completions/mean_length': 104.75, 'completions/min_length': 43.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 46.57143020629883, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.6526471376419067, 'rewards/meter/std': 0.42295220494270325, '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/lexical_plausibility/mean': 0.9645575284957886, 'rewards/lexical_plausibility/std': 0.10024634003639221, 'rewards/judge_quality/mean': 0.6837499737739563, 'rewards/judge_quality/std': 0.3504257798194885, 'rewards/repeat_penalty/mean': 0.9058270454406738, 'rewards/repeat_penalty/std': 0.07923030108213425, 'rewards/near_duplicate_penalty/mean': 0.9540822505950928, 'rewards/near_duplicate_penalty/std': 0.029933732002973557, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9816433191299438, 'rewards/distinct_2/std': 0.03404124453663826, 'rewards/total_composite/mean': 0.47145703434944153, 'rewards/total_composite/std': 0.3599887788295746, 'reward': 0.47145703434944153, 'reward_std': 0.3599887490272522, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13750328123569489, 'sampling/sampling_logp_difference/max': 1.3178296089172363, 'sampling/importance_sampling_ratio/min': 0.267715722322464, 'sampling/importance_sampling_ratio/mean': 1.0084325075149536, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8614685535430908, 'clip_ratio/low_mean': 0.07523171696811914, 'clip_ratio/low_min': 0.07523171696811914, 'clip_ratio/high_mean': 0.05596247687935829, 'clip_ratio/high_max': 0.05596247687935829, 'clip_ratio/region_mean': 0.13119419384747744, 'reward_total_mean': 0.47145703434944153, 'reward_meter_mean': 0.6526471376419067, 'reward_meter_std': 0.42295220494270325, '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_lexical_plausibility_mean': 0.9645575284957886, 'reward_lexical_plausibility_std': 0.10024634003639221, 'reward_repeat_penalty_mean': 0.9058270454406738, 'reward_repeat_penalty_std': 0.07923030108213425, 'reward_near_duplicate_penalty_mean': 0.9540822505950928, 'reward_near_duplicate_penalty_std': 0.029933732002973557, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9816433191299438, 'reward_distinct_2_std': 0.03404124453663826, 'reward_judge_quality_mean': 0.6837499737739563, 'reward_judge_quality_std': 0.3504257798194885, 'reward_total_composite_mean': 0.47145703434944153, 'reward_total_composite_std': 0.3599887788295746, 'epoch': 0.02}

 14%|█▍        | 477/3300 [1:50:34<10:20:30, 13.19s/it]INFO 04-12 12:36:46 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 478/3300 [1:50:47<10:27:43, 13.35s/it]2026-04-12 12:37:00,046 | INFO | train_grpo_train | metrics_logged mode=train step=478

                                                       
{'loss': 0.0476, 'grad_norm': 3.7853403091430664, 'learning_rate': 8.554545454545456e-06, 'num_tokens': 947350.0, 'completions/mean_length': 280.0, 'completions/min_length': 207.0, 'completions/max_length': 334.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 280.0, 'completions/min_terminated_length': 207.0, 'completions/max_terminated_length': 334.0, 'rewards/meter/mean': 0.856927752494812, 'rewards/meter/std': 0.15546144545078278, 'rewards/count_adherence/mean': 0.8055555820465088, 'rewards/count_adherence/std': 0.07856741547584534, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.20000000298023224, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.05588521063327789, 'rewards/repeat_penalty/std': 0.05279583856463432, 'rewards/near_duplicate_penalty/mean': 0.7472454309463501, 'rewards/near_duplicate_penalty/std': 0.06751756370067596, 'rewards/opening_diversity/mean': 0.9114447832107544, 'rewards/opening_diversity/std': 0.09489770978689194, 'rewards/distinct_2/mean': 0.10221026837825775, 'rewards/distinct_2/std': 0.09333431720733643, 'rewards/total_composite/mean': 0.1349012851715088, 'rewards/total_composite/std': 0.09084027260541916, 'reward': 0.1349012851715088, 'reward_std': 0.09084026515483856, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07389570027589798, 'sampling/sampling_logp_difference/max': 2.0485010147094727, 'sampling/importance_sampling_ratio/min': 0.1289280205965042, 'sampling/importance_sampling_ratio/mean': 1.0014927387237549, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5268245153129101, 'clip_ratio/low_mean': 0.048272002255544066, 'clip_ratio/low_min': 0.048272002255544066, 'clip_ratio/high_mean': 0.021325896494090557, 'clip_ratio/high_max': 0.021325896494090557, 'clip_ratio/region_mean': 0.06959789874963462, 'reward_total_mean': 0.1349012851715088, 'reward_meter_mean': 0.856927752494812, 'reward_meter_std': 0.15546144545078278, 'reward_count_adherence_mean': 0.8055555820465088, 'reward_count_adherence_std': 0.07856741547584534, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.05588521063327789, 'reward_repeat_penalty_std': 0.05279583856463432, 'reward_near_duplicate_penalty_mean': 0.7472454309463501, 'reward_near_duplicate_penalty_std': 0.06751756370067596, 'reward_opening_diversity_mean': 0.9114447832107544, 'reward_opening_diversity_std': 0.09489770978689194, 'reward_distinct_2_mean': 0.10221026837825775, 'reward_distinct_2_std': 0.09333431720733643, 'reward_judge_quality_mean': 0.20000000298023224, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.1349012851715088, 'reward_total_composite_std': 0.09084027260541916, 'epoch': 0.02}

 14%|█▍        | 478/3300 [1:50:48<10:27:43, 13.35s/it]INFO 04-12 12:37:00 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 479/3300 [1:50:58<9:41:41, 12.37s/it] 2026-04-12 12:37:10,146 | INFO | train_grpo_train | metrics_logged mode=train step=479

                                                      
{'loss': 0.1001, 'grad_norm': 10.255093574523926, 'learning_rate': 8.551515151515152e-06, 'num_tokens': 949221.0, 'completions/mean_length': 64.875, 'completions/min_length': 51.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.875, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.852449357509613, 'rewards/meter/std': 0.2823321223258972, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9791666269302368, 'rewards/lexical_plausibility/std': 0.0589255727827549, 'rewards/judge_quality/mean': 0.4712499976158142, 'rewards/judge_quality/std': 0.18795421719551086, 'rewards/repeat_penalty/mean': 0.8686116933822632, 'rewards/repeat_penalty/std': 0.19355812668800354, 'rewards/near_duplicate_penalty/mean': 0.9721012115478516, 'rewards/near_duplicate_penalty/std': 0.047310709953308105, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9392306804656982, 'rewards/distinct_2/std': 0.09503848105669022, 'rewards/total_composite/mean': 0.3542850911617279, 'rewards/total_composite/std': 0.24541117250919342, 'reward': 0.3542850911617279, 'reward_std': 0.24541117250919342, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17278730869293213, 'sampling/sampling_logp_difference/max': 1.7193493843078613, 'sampling/importance_sampling_ratio/min': 0.17918269336223602, 'sampling/importance_sampling_ratio/mean': 1.0197672843933105, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3455843776464462, 'clip_ratio/low_mean': 0.1048482395708561, 'clip_ratio/low_min': 0.1048482395708561, 'clip_ratio/high_mean': 0.038804637268185616, 'clip_ratio/high_max': 0.038804637268185616, 'clip_ratio/region_mean': 0.1436528768390417, 'reward_total_mean': 0.3542850911617279, 'reward_meter_mean': 0.852449357509613, 'reward_meter_std': 0.2823321223258972, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9791666269302368, 'reward_lexical_plausibility_std': 0.0589255727827549, 'reward_repeat_penalty_mean': 0.8686116933822632, 'reward_repeat_penalty_std': 0.19355812668800354, 'reward_near_duplicate_penalty_mean': 0.9721012115478516, 'reward_near_duplicate_penalty_std': 0.047310709953308105, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9392306804656982, 'reward_distinct_2_std': 0.09503848105669022, 'reward_judge_quality_mean': 0.4712499976158142, 'reward_judge_quality_std': 0.18795421719551086, 'reward_total_composite_mean': 0.3542850911617279, 'reward_total_composite_std': 0.24541117250919342, 'epoch': 0.02}

 15%|█▍        | 479/3300 [1:50:58<9:41:41, 12.37s/it]INFO 04-12 12:37:10 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 480/3300 [1:51:06<8:46:04, 11.19s/it]2026-04-12 12:37:18,602 | INFO | train_grpo_train | metrics_logged mode=train step=480

                                                      
{'loss': 0.1027, 'grad_norm': 13.397278785705566, 'learning_rate': 8.548484848484849e-06, 'num_tokens': 950981.0, 'completions/mean_length': 34.0, 'completions/min_length': 25.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.0, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.7225872278213501, 'rewards/meter/std': 0.3653312027454376, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7275000214576721, 'rewards/judge_quality/std': 0.3564407229423523, 'rewards/repeat_penalty/mean': 0.533305287361145, 'rewards/repeat_penalty/std': 0.32675471901893616, 'rewards/near_duplicate_penalty/mean': 0.8415569067001343, 'rewards/near_duplicate_penalty/std': 0.3418532907962799, 'rewards/opening_diversity/mean': 0.65625, 'rewards/opening_diversity/std': 0.2651650309562683, 'rewards/distinct_2/mean': 0.7729215621948242, 'rewards/distinct_2/std': 0.38492342829704285, 'rewards/total_composite/mean': 0.4724855422973633, 'rewards/total_composite/std': 0.357770711183548, 'reward': 0.4724855422973633, 'reward_std': 0.357770711183548, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11613781750202179, 'sampling/sampling_logp_difference/max': 1.485879898071289, 'sampling/importance_sampling_ratio/min': 0.22630314528942108, 'sampling/importance_sampling_ratio/mean': 1.0460658073425293, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9398307651281357, 'clip_ratio/low_mean': 0.08029270078986883, 'clip_ratio/low_min': 0.08029270078986883, 'clip_ratio/high_mean': 0.04545029904693365, 'clip_ratio/high_max': 0.04545029904693365, 'clip_ratio/region_mean': 0.12574299983680248, 'reward_total_mean': 0.4724855422973633, 'reward_meter_mean': 0.7225872278213501, 'reward_meter_std': 0.3653312027454376, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.533305287361145, 'reward_repeat_penalty_std': 0.32675471901893616, 'reward_near_duplicate_penalty_mean': 0.8415569067001343, 'reward_near_duplicate_penalty_std': 0.3418532907962799, 'reward_opening_diversity_mean': 0.65625, 'reward_opening_diversity_std': 0.2651650309562683, 'reward_distinct_2_mean': 0.7729215621948242, 'reward_distinct_2_std': 0.38492342829704285, 'reward_judge_quality_mean': 0.7275000214576721, 'reward_judge_quality_std': 0.3564407229423523, 'reward_total_composite_mean': 0.4724855422973633, 'reward_total_composite_std': 0.357770711183548, 'epoch': 0.02}

 15%|█▍        | 480/3300 [1:51:06<8:46:04, 11.19s/it]INFO 04-12 12:37:18 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 481/3300 [1:51:19<9:13:13, 11.77s/it]2026-04-12 12:37:31,719 | INFO | train_grpo_train | metrics_logged mode=train step=481

                                                      
{'loss': -0.0464, 'grad_norm': 6.054765701293945, 'learning_rate': 8.545454545454546e-06, 'num_tokens': 952501.0, 'completions/mean_length': 92.0, 'completions/min_length': 29.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 32.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.7095160484313965, 'rewards/meter/std': 0.4322298467159271, '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/lexical_plausibility/mean': 0.9504179954528809, 'rewards/lexical_plausibility/std': 0.08134390413761139, 'rewards/judge_quality/mean': 0.6974999904632568, 'rewards/judge_quality/std': 0.3388109505176544, 'rewards/repeat_penalty/mean': 0.9867280721664429, 'rewards/repeat_penalty/std': 0.01586657017469406, 'rewards/near_duplicate_penalty/mean': 0.9872534871101379, 'rewards/near_duplicate_penalty/std': 0.016274144873023033, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9994745850563049, 'rewards/distinct_2/std': 0.0014861400704830885, 'rewards/total_composite/mean': 0.5951536893844604, 'rewards/total_composite/std': 0.39523252844810486, 'reward': 0.5951536893844604, 'reward_std': 0.39523252844810486, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15185675024986267, 'sampling/sampling_logp_difference/max': 1.6311500072479248, 'sampling/importance_sampling_ratio/min': 0.195704385638237, 'sampling/importance_sampling_ratio/mean': 1.0125807523727417, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8218754008412361, 'clip_ratio/low_mean': 0.04642857238650322, 'clip_ratio/low_min': 0.04642857238650322, 'clip_ratio/high_mean': 0.07045293226838112, 'clip_ratio/high_max': 0.07045293226838112, 'clip_ratio/region_mean': 0.11688150465488434, 'reward_total_mean': 0.5951536893844604, 'reward_meter_mean': 0.7095160484313965, 'reward_meter_std': 0.4322298467159271, '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_lexical_plausibility_mean': 0.9504179954528809, 'reward_lexical_plausibility_std': 0.08134390413761139, 'reward_repeat_penalty_mean': 0.9867280721664429, 'reward_repeat_penalty_std': 0.01586657017469406, 'reward_near_duplicate_penalty_mean': 0.9872534871101379, 'reward_near_duplicate_penalty_std': 0.016274144873023033, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9994745850563049, 'reward_distinct_2_std': 0.0014861400704830885, 'reward_judge_quality_mean': 0.6974999904632568, 'reward_judge_quality_std': 0.3388109505176544, 'reward_total_composite_mean': 0.5951536893844604, 'reward_total_composite_std': 0.39523252844810486, 'epoch': 0.02}

 15%|█▍        | 481/3300 [1:51:19<9:13:13, 11.77s/it]INFO 04-12 12:37:31 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 482/3300 [1:51:29<8:40:45, 11.09s/it]2026-04-12 12:37:41,207 | INFO | train_grpo_train | metrics_logged mode=train step=482

                                                      
{'loss': 0.0608, 'grad_norm': 15.708717346191406, 'learning_rate': 8.542424242424243e-06, 'num_tokens': 954153.0, 'completions/mean_length': 46.5, 'completions/min_length': 42.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.5, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.8546483516693115, 'rewards/meter/std': 0.30196648836135864, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7512500286102295, 'rewards/judge_quality/std': 0.25542333722114563, 'rewards/repeat_penalty/mean': 0.985704779624939, 'rewards/repeat_penalty/std': 0.014976373873651028, 'rewards/near_duplicate_penalty/mean': 0.985704779624939, 'rewards/near_duplicate_penalty/std': 0.014976373873651028, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6230180859565735, 'rewards/total_composite/std': 0.3138609528541565, 'reward': 0.6230180859565735, 'reward_std': 0.3138609230518341, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14824171364307404, 'sampling/sampling_logp_difference/max': 2.219102382659912, 'sampling/importance_sampling_ratio/min': 0.10870664566755295, 'sampling/importance_sampling_ratio/mean': 1.028839111328125, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9098098948597908, 'clip_ratio/low_mean': 0.04341503418982029, 'clip_ratio/low_min': 0.04341503418982029, 'clip_ratio/high_mean': 0.07458710484206676, 'clip_ratio/high_max': 0.07458710484206676, 'clip_ratio/region_mean': 0.11800213903188705, 'reward_total_mean': 0.6230180859565735, 'reward_meter_mean': 0.8546483516693115, 'reward_meter_std': 0.30196648836135864, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.985704779624939, 'reward_repeat_penalty_std': 0.014976373873651028, 'reward_near_duplicate_penalty_mean': 0.985704779624939, 'reward_near_duplicate_penalty_std': 0.014976373873651028, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7512500286102295, 'reward_judge_quality_std': 0.25542333722114563, 'reward_total_composite_mean': 0.6230180859565735, 'reward_total_composite_std': 0.3138609528541565, 'epoch': 0.02}

 15%|█▍        | 482/3300 [1:51:29<8:40:45, 11.09s/it]INFO 04-12 12:37:41 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 483/3300 [1:51:37<8:05:22, 10.34s/it]2026-04-12 12:37:49,795 | INFO | train_grpo_train | metrics_logged mode=train step=483

                                                      
{'loss': 0.0853, 'grad_norm': 14.387272834777832, 'learning_rate': 8.539393939393939e-06, 'num_tokens': 955707.0, 'completions/mean_length': 42.25, 'completions/min_length': 37.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.25, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.6656347513198853, 'rewards/meter/std': 0.3080311417579651, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6887500286102295, 'rewards/judge_quality/std': 0.31903597712516785, 'rewards/repeat_penalty/mean': 0.9446030259132385, 'rewards/repeat_penalty/std': 0.05046504735946655, 'rewards/near_duplicate_penalty/mean': 0.9620427489280701, 'rewards/near_duplicate_penalty/std': 0.03469841182231903, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9816849827766418, 'rewards/distinct_2/std': 0.03391282260417938, 'rewards/total_composite/mean': 0.3735358715057373, 'rewards/total_composite/std': 0.3373759388923645, 'reward': 0.3735358715057373, 'reward_std': 0.3373759090900421, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15325802564620972, 'sampling/sampling_logp_difference/max': 1.6850972175598145, 'sampling/importance_sampling_ratio/min': 0.18542641401290894, 'sampling/importance_sampling_ratio/mean': 1.0021201372146606, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0205260142683983, 'clip_ratio/low_mean': 0.04210150986909866, 'clip_ratio/low_min': 0.04210150986909866, 'clip_ratio/high_mean': 0.08551118616014719, 'clip_ratio/high_max': 0.08551118616014719, 'clip_ratio/region_mean': 0.12761269602924585, 'reward_total_mean': 0.3735358715057373, 'reward_meter_mean': 0.6656347513198853, 'reward_meter_std': 0.3080311417579651, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9446030259132385, 'reward_repeat_penalty_std': 0.05046504735946655, 'reward_near_duplicate_penalty_mean': 0.9620427489280701, 'reward_near_duplicate_penalty_std': 0.03469841182231903, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9816849827766418, 'reward_distinct_2_std': 0.03391282260417938, 'reward_judge_quality_mean': 0.6887500286102295, 'reward_judge_quality_std': 0.31903597712516785, 'reward_total_composite_mean': 0.3735358715057373, 'reward_total_composite_std': 0.3373759388923645, 'epoch': 0.02}

 15%|█▍        | 483/3300 [1:51:37<8:05:22, 10.34s/it]INFO 04-12 12:37:50 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 484/3300 [1:51:46<7:43:18,  9.87s/it]2026-04-12 12:37:58,582 | INFO | train_grpo_train | metrics_logged mode=train step=484

                                                      
{'loss': 0.1009, 'grad_norm': 17.388076782226562, 'learning_rate': 8.536363636363636e-06, 'num_tokens': 957339.0, 'completions/mean_length': 56.0, 'completions/min_length': 49.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.0, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.659206748008728, 'rewards/meter/std': 0.35508015751838684, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7887499928474426, 'rewards/judge_quality/std': 0.27115821838378906, 'rewards/repeat_penalty/mean': 0.9011809229850769, 'rewards/repeat_penalty/std': 0.22922390699386597, 'rewards/near_duplicate_penalty/mean': 0.9623961448669434, 'rewards/near_duplicate_penalty/std': 0.07599963992834091, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9409211277961731, 'rewards/distinct_2/std': 0.14684276282787323, 'rewards/total_composite/mean': 0.5697370171546936, 'rewards/total_composite/std': 0.3841080665588379, 'reward': 0.5697370171546936, 'reward_std': 0.3841080665588379, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1469540297985077, 'sampling/sampling_logp_difference/max': 1.8108291625976562, 'sampling/importance_sampling_ratio/min': 0.163518488407135, 'sampling/importance_sampling_ratio/mean': 0.9772599339485168, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7122868299484253, 'clip_ratio/low_mean': 0.03883105143904686, 'clip_ratio/low_min': 0.03883105143904686, 'clip_ratio/high_mean': 0.07303191209211946, 'clip_ratio/high_max': 0.07303191209211946, 'clip_ratio/region_mean': 0.11186296353116632, 'reward_total_mean': 0.5697370171546936, 'reward_meter_mean': 0.659206748008728, 'reward_meter_std': 0.35508015751838684, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9011809229850769, 'reward_repeat_penalty_std': 0.22922390699386597, 'reward_near_duplicate_penalty_mean': 0.9623961448669434, 'reward_near_duplicate_penalty_std': 0.07599963992834091, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9409211277961731, 'reward_distinct_2_std': 0.14684276282787323, 'reward_judge_quality_mean': 0.7887499928474426, 'reward_judge_quality_std': 0.27115821838378906, 'reward_total_composite_mean': 0.5697370171546936, 'reward_total_composite_std': 0.3841080665588379, 'epoch': 0.02}

 15%|█▍        | 484/3300 [1:51:46<7:43:18,  9.87s/it]INFO 04-12 12:37:58 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 485/3300 [1:51:56<7:46:51,  9.95s/it]2026-04-12 12:38:08,713 | INFO | train_grpo_train | metrics_logged mode=train step=485

                                                      
{'loss': 0.0296, 'grad_norm': 14.966595649719238, 'learning_rate': 8.533333333333335e-06, 'num_tokens': 959394.0, 'completions/mean_length': 86.875, 'completions/min_length': 77.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 86.875, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.3842068910598755, 'rewards/meter/std': 0.3022935390472412, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6387499570846558, 'rewards/judge_quality/std': 0.30884289741516113, 'rewards/repeat_penalty/mean': 0.9969550967216492, 'rewards/repeat_penalty/std': 0.005105931777507067, 'rewards/near_duplicate_penalty/mean': 0.9969550967216492, 'rewards/near_duplicate_penalty/std': 0.005105931777507067, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.21432170271873474, 'rewards/total_composite/std': 0.18373489379882812, 'reward': 0.21432170271873474, 'reward_std': 0.18373489379882812, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18125201761722565, 'sampling/sampling_logp_difference/max': 3.1263813972473145, 'sampling/importance_sampling_ratio/min': 0.0438762828707695, 'sampling/importance_sampling_ratio/mean': 0.9934475421905518, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8047947511076927, 'clip_ratio/low_mean': 0.09030736144632101, 'clip_ratio/low_min': 0.09030736144632101, 'clip_ratio/high_mean': 0.06894026137888432, 'clip_ratio/high_max': 0.06894026137888432, 'clip_ratio/region_mean': 0.15924762282520533, 'reward_total_mean': 0.21432170271873474, 'reward_meter_mean': 0.3842068910598755, 'reward_meter_std': 0.3022935390472412, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9969550967216492, 'reward_repeat_penalty_std': 0.005105931777507067, 'reward_near_duplicate_penalty_mean': 0.9969550967216492, 'reward_near_duplicate_penalty_std': 0.005105931777507067, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6387499570846558, 'reward_judge_quality_std': 0.30884289741516113, 'reward_total_composite_mean': 0.21432170271873474, 'reward_total_composite_std': 0.18373489379882812, 'epoch': 0.02}

 15%|█▍        | 485/3300 [1:51:56<7:46:51,  9.95s/it]INFO 04-12 12:38:08 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 486/3300 [1:52:05<7:31:51,  9.63s/it]2026-04-12 12:38:17,611 | INFO | train_grpo_train | metrics_logged mode=train step=486

                                                      
{'loss': -0.0064, 'grad_norm': 12.775290489196777, 'learning_rate': 8.53030303030303e-06, 'num_tokens': 961187.0, 'completions/mean_length': 53.125, 'completions/min_length': 48.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.125, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.984459400177002, 'rewards/meter/std': 0.01069045253098011, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.4625000059604645, 'rewards/judge_quality/std': 0.15526476502418518, 'rewards/repeat_penalty/mean': 0.959327220916748, 'rewards/repeat_penalty/std': 0.041978754103183746, 'rewards/near_duplicate_penalty/mean': 0.9671092629432678, 'rewards/near_duplicate_penalty/std': 0.02511446923017502, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9916387796401978, 'rewards/distinct_2/std': 0.0236490610986948, 'rewards/total_composite/mean': 0.4558395743370056, 'rewards/total_composite/std': 0.15499630570411682, 'reward': 0.4558395743370056, 'reward_std': 0.15499630570411682, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1525525599718094, 'sampling/sampling_logp_difference/max': 1.3094673156738281, 'sampling/importance_sampling_ratio/min': 0.2699638307094574, 'sampling/importance_sampling_ratio/mean': 1.0213919878005981, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.192850761115551, 'clip_ratio/low_mean': 0.09158541914075613, 'clip_ratio/low_min': 0.09158541914075613, 'clip_ratio/high_mean': 0.03660714440047741, 'clip_ratio/high_max': 0.03660714440047741, 'clip_ratio/region_mean': 0.12819256354123354, 'reward_total_mean': 0.4558395743370056, 'reward_meter_mean': 0.984459400177002, 'reward_meter_std': 0.01069045253098011, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.959327220916748, 'reward_repeat_penalty_std': 0.041978754103183746, 'reward_near_duplicate_penalty_mean': 0.9671092629432678, 'reward_near_duplicate_penalty_std': 0.02511446923017502, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9916387796401978, 'reward_distinct_2_std': 0.0236490610986948, 'reward_judge_quality_mean': 0.4625000059604645, 'reward_judge_quality_std': 0.15526476502418518, 'reward_total_composite_mean': 0.4558395743370056, 'reward_total_composite_std': 0.15499630570411682, 'epoch': 0.02}

 15%|█▍        | 486/3300 [1:52:05<7:31:51,  9.63s/it]INFO 04-12 12:38:17 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 487/3300 [1:52:15<7:33:33,  9.67s/it]2026-04-12 12:38:27,379 | INFO | train_grpo_train | metrics_logged mode=train step=487

                                                      
{'loss': 0.0688, 'grad_norm': 11.830143928527832, 'learning_rate': 8.527272727272728e-06, 'num_tokens': 963008.0, 'completions/mean_length': 66.625, 'completions/min_length': 57.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.625, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.8934787511825562, 'rewards/meter/std': 0.23271259665489197, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4375, 'rewards/judge_quality/std': 0.13562028110027313, 'rewards/repeat_penalty/mean': 0.5014407634735107, 'rewards/repeat_penalty/std': 0.32434895634651184, 'rewards/near_duplicate_penalty/mean': 0.8915968537330627, 'rewards/near_duplicate_penalty/std': 0.08636351674795151, 'rewards/opening_diversity/mean': 0.84375, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.69657963514328, 'rewards/distinct_2/std': 0.23540747165679932, 'rewards/total_composite/mean': 0.390258252620697, 'rewards/total_composite/std': 0.16791626811027527, 'reward': 0.390258252620697, 'reward_std': 0.16791626811027527, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1091838851571083, 'sampling/sampling_logp_difference/max': 1.3491935729980469, 'sampling/importance_sampling_ratio/min': 0.25944939255714417, 'sampling/importance_sampling_ratio/mean': 1.0011334419250488, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7564239725470543, 'clip_ratio/low_mean': 0.05169323552399874, 'clip_ratio/low_min': 0.05169323552399874, 'clip_ratio/high_mean': 0.042048532515764236, 'clip_ratio/high_max': 0.042048532515764236, 'clip_ratio/region_mean': 0.09374176803976297, 'reward_total_mean': 0.390258252620697, 'reward_meter_mean': 0.8934787511825562, 'reward_meter_std': 0.23271259665489197, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.5014407634735107, 'reward_repeat_penalty_std': 0.32434895634651184, 'reward_near_duplicate_penalty_mean': 0.8915968537330627, 'reward_near_duplicate_penalty_std': 0.08636351674795151, 'reward_opening_diversity_mean': 0.84375, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.69657963514328, 'reward_distinct_2_std': 0.23540747165679932, 'reward_judge_quality_mean': 0.4375, 'reward_judge_quality_std': 0.13562028110027313, 'reward_total_composite_mean': 0.390258252620697, 'reward_total_composite_std': 0.16791626811027527, 'epoch': 0.02}

 15%|█▍        | 487/3300 [1:52:15<7:33:33,  9.67s/it]INFO 04-12 12:38:27 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 488/3300 [1:52:26<7:54:55, 10.13s/it]2026-04-12 12:38:38,587 | INFO | train_grpo_train | metrics_logged mode=train step=488

                                                      
{'loss': 0.062, 'grad_norm': 8.0690336227417, 'learning_rate': 8.524242424242425e-06, 'num_tokens': 965669.0, 'completions/mean_length': 144.625, 'completions/min_length': 122.0, 'completions/max_length': 168.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 144.625, 'completions/min_terminated_length': 122.0, 'completions/max_terminated_length': 168.0, 'rewards/meter/mean': 0.8518550992012024, 'rewards/meter/std': 0.2522626221179962, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.5154203176498413, 'rewards/repeat_penalty/std': 0.4040014445781708, 'rewards/near_duplicate_penalty/mean': 0.8473269939422607, 'rewards/near_duplicate_penalty/std': 0.16172580420970917, 'rewards/opening_diversity/mean': 0.8656250238418579, 'rewards/opening_diversity/std': 0.24601738154888153, 'rewards/distinct_2/mean': 0.606722354888916, 'rewards/distinct_2/std': 0.3611520230770111, 'rewards/total_composite/mean': 0.29777130484580994, 'rewards/total_composite/std': 0.1329665184020996, 'reward': 0.29777130484580994, 'reward_std': 0.13296650350093842, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11718686670064926, 'sampling/sampling_logp_difference/max': 2.008418083190918, 'sampling/importance_sampling_ratio/min': 0.1342008113861084, 'sampling/importance_sampling_ratio/mean': 1.0089386701583862, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7607554048299789, 'clip_ratio/low_mean': 0.042041667737066746, 'clip_ratio/low_min': 0.042041667737066746, 'clip_ratio/high_mean': 0.07441166881471872, 'clip_ratio/high_max': 0.07441166881471872, 'clip_ratio/region_mean': 0.11645333655178547, 'reward_total_mean': 0.29777130484580994, 'reward_meter_mean': 0.8518550992012024, 'reward_meter_std': 0.2522626221179962, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.09258200973272324, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.5154203176498413, 'reward_repeat_penalty_std': 0.4040014445781708, 'reward_near_duplicate_penalty_mean': 0.8473269939422607, 'reward_near_duplicate_penalty_std': 0.16172580420970917, 'reward_opening_diversity_mean': 0.8656250238418579, 'reward_opening_diversity_std': 0.24601738154888153, 'reward_distinct_2_mean': 0.606722354888916, 'reward_distinct_2_std': 0.3611520230770111, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.29777130484580994, 'reward_total_composite_std': 0.1329665184020996, 'epoch': 0.02}

 15%|█▍        | 488/3300 [1:52:26<7:54:55, 10.13s/it]INFO 04-12 12:38:38 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 489/3300 [1:52:36<7:48:45, 10.01s/it]2026-04-12 12:38:48,295 | INFO | train_grpo_train | metrics_logged mode=train step=489

                                                      
{'loss': 0.0272, 'grad_norm': 10.402389526367188, 'learning_rate': 8.521212121212123e-06, 'num_tokens': 967710.0, 'completions/mean_length': 81.125, 'completions/min_length': 62.0, 'completions/max_length': 109.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 81.125, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 109.0, 'rewards/meter/mean': 0.3735775947570801, 'rewards/meter/std': 0.4236741065979004, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.955470621585846, 'rewards/lexical_plausibility/std': 0.08526492863893509, 'rewards/judge_quality/mean': 0.44624999165534973, 'rewards/judge_quality/std': 0.19668231904506683, 'rewards/repeat_penalty/mean': 0.6183725595474243, 'rewards/repeat_penalty/std': 0.318236380815506, 'rewards/near_duplicate_penalty/mean': 0.8775686025619507, 'rewards/near_duplicate_penalty/std': 0.10264322906732559, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.7508325576782227, 'rewards/distinct_2/std': 0.23671065270900726, 'rewards/total_composite/mean': 0.13460922241210938, 'rewards/total_composite/std': 0.16504879295825958, 'reward': 0.13460922241210938, 'reward_std': 0.1650487780570984, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14535221457481384, 'sampling/sampling_logp_difference/max': 1.6159615516662598, 'sampling/importance_sampling_ratio/min': 0.19869951903820038, 'sampling/importance_sampling_ratio/mean': 1.0160772800445557, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8481792211532593, 'clip_ratio/low_mean': 0.07692583510652184, 'clip_ratio/low_min': 0.07692583510652184, 'clip_ratio/high_mean': 0.06324087642133236, 'clip_ratio/high_max': 0.06324087642133236, 'clip_ratio/region_mean': 0.1401667115278542, 'reward_total_mean': 0.13460922241210938, 'reward_meter_mean': 0.3735775947570801, 'reward_meter_std': 0.4236741065979004, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.955470621585846, 'reward_lexical_plausibility_std': 0.08526492863893509, 'reward_repeat_penalty_mean': 0.6183725595474243, 'reward_repeat_penalty_std': 0.318236380815506, 'reward_near_duplicate_penalty_mean': 0.8775686025619507, 'reward_near_duplicate_penalty_std': 0.10264322906732559, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.7508325576782227, 'reward_distinct_2_std': 0.23671065270900726, 'reward_judge_quality_mean': 0.44624999165534973, 'reward_judge_quality_std': 0.19668231904506683, 'reward_total_composite_mean': 0.13460922241210938, 'reward_total_composite_std': 0.16504879295825958, 'epoch': 0.02}

 15%|█▍        | 489/3300 [1:52:36<7:48:45, 10.01s/it]INFO 04-12 12:38:48 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 490/3300 [1:52:47<8:02:00, 10.29s/it]2026-04-12 12:38:59,251 | INFO | train_grpo_train | metrics_logged mode=train step=490

                                                      
{'loss': 0.0677, 'grad_norm': 6.43690299987793, 'learning_rate': 8.518181818181818e-06, 'num_tokens': 970834.0, 'completions/mean_length': 164.5, 'completions/min_length': 149.0, 'completions/max_length': 198.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 164.5, 'completions/min_terminated_length': 149.0, 'completions/max_terminated_length': 198.0, 'rewards/meter/mean': 0.561223030090332, 'rewards/meter/std': 0.4072257876396179, 'rewards/count_adherence/mean': 0.8958333134651184, 'rewards/count_adherence/std': 0.08625820279121399, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.23750001192092896, 'rewards/judge_quality/std': 0.12464234977960587, 'rewards/repeat_penalty/mean': 0.4765090346336365, 'rewards/repeat_penalty/std': 0.38373279571533203, 'rewards/near_duplicate_penalty/mean': 0.8363927602767944, 'rewards/near_duplicate_penalty/std': 0.1408124417066574, 'rewards/opening_diversity/mean': 0.9299107193946838, 'rewards/opening_diversity/std': 0.14020416140556335, 'rewards/distinct_2/mean': 0.5466804504394531, 'rewards/distinct_2/std': 0.3601234257221222, 'rewards/total_composite/mean': 0.12919865548610687, 'rewards/total_composite/std': 0.1509620100259781, 'reward': 0.12919865548610687, 'reward_std': 0.1509620100259781, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11495618522167206, 'sampling/sampling_logp_difference/max': 2.230985164642334, 'sampling/importance_sampling_ratio/min': 0.10742254555225372, 'sampling/importance_sampling_ratio/mean': 1.0070220232009888, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.661452479660511, 'clip_ratio/low_mean': 0.06424165517091751, 'clip_ratio/low_min': 0.06424165517091751, 'clip_ratio/high_mean': 0.05556938424706459, 'clip_ratio/high_max': 0.05556938424706459, 'clip_ratio/region_mean': 0.1198110394179821, 'reward_total_mean': 0.12919865548610687, 'reward_meter_mean': 0.561223030090332, 'reward_meter_std': 0.4072257876396179, 'reward_count_adherence_mean': 0.8958333134651184, 'reward_count_adherence_std': 0.08625820279121399, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.4765090346336365, 'reward_repeat_penalty_std': 0.38373279571533203, 'reward_near_duplicate_penalty_mean': 0.8363927602767944, 'reward_near_duplicate_penalty_std': 0.1408124417066574, 'reward_opening_diversity_mean': 0.9299107193946838, 'reward_opening_diversity_std': 0.14020416140556335, 'reward_distinct_2_mean': 0.5466804504394531, 'reward_distinct_2_std': 0.3601234257221222, 'reward_judge_quality_mean': 0.23750001192092896, 'reward_judge_quality_std': 0.12464234977960587, 'reward_total_composite_mean': 0.12919865548610687, 'reward_total_composite_std': 0.1509620100259781, 'epoch': 0.02}

 15%|█▍        | 490/3300 [1:52:47<8:02:00, 10.29s/it]INFO 04-12 12:38:59 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 491/3300 [1:53:02<9:09:30, 11.74s/it]2026-04-12 12:39:14,359 | INFO | train_grpo_train | metrics_logged mode=train step=491

                                                      
{'loss': -0.0577, 'grad_norm': 6.091986179351807, 'learning_rate': 8.515151515151517e-06, 'num_tokens': 972321.0, 'completions/mean_length': 102.875, 'completions/min_length': 41.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 44.42857360839844, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.6568453907966614, 'rewards/meter/std': 0.423616498708725, '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/lexical_plausibility/mean': 0.9624043107032776, 'rewards/lexical_plausibility/std': 0.10633666813373566, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.155264750123024, 'rewards/repeat_penalty/mean': 0.7982226014137268, 'rewards/repeat_penalty/std': 0.2836403250694275, 'rewards/near_duplicate_penalty/mean': 0.9203413128852844, 'rewards/near_duplicate_penalty/std': 0.08349907398223877, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9157509207725525, 'rewards/distinct_2/std': 0.13582074642181396, 'rewards/total_composite/mean': 0.19255058467388153, 'rewards/total_composite/std': 0.17068003118038177, 'reward': 0.19255058467388153, 'reward_std': 0.17068003118038177, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1475788652896881, 'sampling/sampling_logp_difference/max': 1.4643404483795166, 'sampling/importance_sampling_ratio/min': 0.2312304526567459, 'sampling/importance_sampling_ratio/mean': 1.0223606824874878, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6598791852593422, 'clip_ratio/low_mean': 0.0766980443149805, 'clip_ratio/low_min': 0.0766980443149805, 'clip_ratio/high_mean': 0.05172917549498379, 'clip_ratio/high_max': 0.05172917549498379, 'clip_ratio/region_mean': 0.1284272198099643, 'reward_total_mean': 0.19255058467388153, 'reward_meter_mean': 0.6568453907966614, 'reward_meter_std': 0.423616498708725, '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_lexical_plausibility_mean': 0.9624043107032776, 'reward_lexical_plausibility_std': 0.10633666813373566, 'reward_repeat_penalty_mean': 0.7982226014137268, 'reward_repeat_penalty_std': 0.2836403250694275, 'reward_near_duplicate_penalty_mean': 0.9203413128852844, 'reward_near_duplicate_penalty_std': 0.08349907398223877, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9157509207725525, 'reward_distinct_2_std': 0.13582074642181396, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.155264750123024, 'reward_total_composite_mean': 0.19255058467388153, 'reward_total_composite_std': 0.17068003118038177, 'epoch': 0.02}

 15%|█▍        | 491/3300 [1:53:02<9:09:30, 11.74s/it]INFO 04-12 12:39:14 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 492/3300 [1:53:19<10:20:23, 13.26s/it]2026-04-12 12:39:31,164 | INFO | train_grpo_train | metrics_logged mode=train step=492

                                                       
{'loss': -0.056, 'grad_norm': 5.773955821990967, 'learning_rate': 8.512121212121213e-06, 'num_tokens': 974065.0, 'completions/mean_length': 102.0, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 43.42857360839844, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.3973095118999481, 'rewards/meter/std': 0.2902991771697998, '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/lexical_plausibility/mean': 0.9618542194366455, 'rewards/lexical_plausibility/std': 0.10789252072572708, 'rewards/judge_quality/mean': 0.4087499976158142, 'rewards/judge_quality/std': 0.24121345579624176, 'rewards/repeat_penalty/mean': 0.9809025526046753, 'rewards/repeat_penalty/std': 0.012960494495928288, 'rewards/near_duplicate_penalty/mean': 0.9834362268447876, 'rewards/near_duplicate_penalty/std': 0.010643698275089264, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9974221587181091, 'rewards/distinct_2/std': 0.007291257381439209, 'rewards/total_composite/mean': 0.19527246057987213, 'rewards/total_composite/std': 0.2146981805562973, 'reward': 0.19527246057987213, 'reward_std': 0.2146981656551361, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11211026459932327, 'sampling/sampling_logp_difference/max': 1.768599033355713, 'sampling/importance_sampling_ratio/min': 0.17057178914546967, 'sampling/importance_sampling_ratio/mean': 1.041520595550537, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6453075483441353, 'clip_ratio/low_mean': 0.06031831633299589, 'clip_ratio/low_min': 0.06031831633299589, 'clip_ratio/high_mean': 0.03870355058461428, 'clip_ratio/high_max': 0.03870355058461428, 'clip_ratio/region_mean': 0.09902186691761017, 'reward_total_mean': 0.19527246057987213, 'reward_meter_mean': 0.3973095118999481, 'reward_meter_std': 0.2902991771697998, '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_lexical_plausibility_mean': 0.9618542194366455, 'reward_lexical_plausibility_std': 0.10789252072572708, 'reward_repeat_penalty_mean': 0.9809025526046753, 'reward_repeat_penalty_std': 0.012960494495928288, 'reward_near_duplicate_penalty_mean': 0.9834362268447876, 'reward_near_duplicate_penalty_std': 0.010643698275089264, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9974221587181091, 'reward_distinct_2_std': 0.007291257381439209, 'reward_judge_quality_mean': 0.4087499976158142, 'reward_judge_quality_std': 0.24121345579624176, 'reward_total_composite_mean': 0.19527246057987213, 'reward_total_composite_std': 0.2146981805562973, 'epoch': 0.02}

 15%|█▍        | 492/3300 [1:53:19<10:20:23, 13.26s/it]INFO 04-12 12:39:31 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 493/3300 [1:53:34<10:48:03, 13.85s/it]2026-04-12 12:39:46,403 | INFO | train_grpo_train | metrics_logged mode=train step=493

                                                       
{'loss': -0.1466, 'grad_norm': 4.135079860687256, 'learning_rate': 8.50909090909091e-06, 'num_tokens': 976415.0, 'completions/mean_length': 150.75, 'completions/min_length': 88.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 99.14286041259766, 'completions/min_terminated_length': 88.0, 'completions/max_terminated_length': 125.0, 'rewards/meter/mean': 0.927754282951355, 'rewards/meter/std': 0.14324016869068146, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/repeat_penalty/mean': 0.787428617477417, 'rewards/repeat_penalty/std': 0.27102622389793396, 'rewards/near_duplicate_penalty/mean': 0.9316419363021851, 'rewards/near_duplicate_penalty/std': 0.0827101320028305, 'rewards/opening_diversity/mean': 0.9765625, 'rewards/opening_diversity/std': 0.032346826046705246, 'rewards/distinct_2/mean': 0.8527272343635559, 'rewards/distinct_2/std': 0.23050810396671295, 'rewards/total_composite/mean': 0.29784029722213745, 'rewards/total_composite/std': 0.11790855973958969, 'reward': 0.29784029722213745, 'reward_std': 0.11790855973958969, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12336722016334534, 'sampling/sampling_logp_difference/max': 1.796236515045166, 'sampling/importance_sampling_ratio/min': 0.165922150015831, 'sampling/importance_sampling_ratio/mean': 1.0029431581497192, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5936805307865143, 'clip_ratio/low_mean': 0.009943181648850441, 'clip_ratio/low_min': 0.009943181648850441, 'clip_ratio/high_mean': 0.09094352554529905, 'clip_ratio/high_max': 0.09094352554529905, 'clip_ratio/region_mean': 0.1008867071941495, 'reward_total_mean': 0.29784029722213745, 'reward_meter_mean': 0.927754282951355, 'reward_meter_std': 0.14324016869068146, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.787428617477417, 'reward_repeat_penalty_std': 0.27102622389793396, 'reward_near_duplicate_penalty_mean': 0.9316419363021851, 'reward_near_duplicate_penalty_std': 0.0827101320028305, 'reward_opening_diversity_mean': 0.9765625, 'reward_opening_diversity_std': 0.032346826046705246, 'reward_distinct_2_mean': 0.8527272343635559, 'reward_distinct_2_std': 0.23050810396671295, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.29784029722213745, 'reward_total_composite_std': 0.11790855973958969, 'epoch': 0.02}

 15%|█▍        | 493/3300 [1:53:34<10:48:03, 13.85s/it]INFO 04-12 12:39:46 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 494/3300 [1:53:48<10:57:41, 14.06s/it]2026-04-12 12:40:00,958 | INFO | train_grpo_train | metrics_logged mode=train step=494

                                                       
{'loss': -0.1774, 'grad_norm': 2.5734379291534424, 'learning_rate': 8.506060606060607e-06, 'num_tokens': 979019.0, 'completions/mean_length': 192.5, 'completions/min_length': 123.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 146.85714721679688, 'completions/min_terminated_length': 123.0, 'completions/max_terminated_length': 165.0, 'rewards/meter/mean': 0.8972945809364319, 'rewards/meter/std': 0.24541336297988892, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9713812470436096, 'rewards/lexical_plausibility/std': 0.08094610273838043, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.20878174901008606, 'rewards/repeat_penalty/std': 0.32008856534957886, 'rewards/near_duplicate_penalty/mean': 0.7826398611068726, 'rewards/near_duplicate_penalty/std': 0.09742272645235062, 'rewards/opening_diversity/mean': 0.7541667222976685, 'rewards/opening_diversity/std': 0.16201850771903992, 'rewards/distinct_2/mean': 0.3760785162448883, 'rewards/distinct_2/std': 0.3009520471096039, 'rewards/total_composite/mean': 0.25990408658981323, 'rewards/total_composite/std': 0.12562499940395355, 'reward': 0.25990408658981323, 'reward_std': 0.12562499940395355, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08453472703695297, 'sampling/sampling_logp_difference/max': 1.376196265220642, 'sampling/importance_sampling_ratio/min': 0.252537339925766, 'sampling/importance_sampling_ratio/mean': 1.0119431018829346, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6304626390337944, 'clip_ratio/low_mean': 0.005303030367940664, 'clip_ratio/low_min': 0.005303030367940664, 'clip_ratio/high_mean': 0.06734117865562439, 'clip_ratio/high_max': 0.06734117865562439, 'clip_ratio/region_mean': 0.07264420902356505, 'reward_total_mean': 0.25990408658981323, 'reward_meter_mean': 0.8972945809364319, 'reward_meter_std': 0.24541336297988892, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.09258200973272324, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9713812470436096, 'reward_lexical_plausibility_std': 0.08094610273838043, 'reward_repeat_penalty_mean': 0.20878174901008606, 'reward_repeat_penalty_std': 0.32008856534957886, 'reward_near_duplicate_penalty_mean': 0.7826398611068726, 'reward_near_duplicate_penalty_std': 0.09742272645235062, 'reward_opening_diversity_mean': 0.7541667222976685, 'reward_opening_diversity_std': 0.16201850771903992, 'reward_distinct_2_mean': 0.3760785162448883, 'reward_distinct_2_std': 0.3009520471096039, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.25990408658981323, 'reward_total_composite_std': 0.12562499940395355, 'epoch': 0.02}

 15%|█▍        | 494/3300 [1:53:48<10:57:41, 14.06s/it]INFO 04-12 12:40:01 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 495/3300 [1:54:02<10:48:46, 13.88s/it]2026-04-12 12:40:14,411 | INFO | train_grpo_train | metrics_logged mode=train step=495

                                                       
{'loss': -0.0679, 'grad_norm': 5.250760078430176, 'learning_rate': 8.503030303030304e-06, 'num_tokens': 980711.0, 'completions/mean_length': 107.5, 'completions/min_length': 46.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 49.71428680419922, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.7184078693389893, 'rewards/meter/std': 0.38564369082450867, '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/lexical_plausibility/mean': 0.9569368958473206, 'rewards/lexical_plausibility/std': 0.12180086970329285, 'rewards/judge_quality/mean': 0.4087499976158142, 'rewards/judge_quality/std': 0.27445724606513977, 'rewards/repeat_penalty/mean': 0.9554755091667175, 'rewards/repeat_penalty/std': 0.03760902211070061, 'rewards/near_duplicate_penalty/mean': 0.9656516313552856, 'rewards/near_duplicate_penalty/std': 0.02569839358329773, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9893856048583984, 'rewards/distinct_2/std': 0.025639133527874947, 'rewards/total_composite/mean': 0.3346109986305237, 'rewards/total_composite/std': 0.28311118483543396, 'reward': 0.3346109986305237, 'reward_std': 0.2831111550331116, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1423669308423996, 'sampling/sampling_logp_difference/max': 1.532524585723877, 'sampling/importance_sampling_ratio/min': 0.21598969399929047, 'sampling/importance_sampling_ratio/mean': 1.0120123624801636, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7360295429825783, 'clip_ratio/low_mean': 0.04146173573099077, 'clip_ratio/low_min': 0.04146173573099077, 'clip_ratio/high_mean': 0.06527573429048061, 'clip_ratio/high_max': 0.06527573429048061, 'clip_ratio/region_mean': 0.10673747002147138, 'reward_total_mean': 0.3346109986305237, 'reward_meter_mean': 0.7184078693389893, 'reward_meter_std': 0.38564369082450867, '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_lexical_plausibility_mean': 0.9569368958473206, 'reward_lexical_plausibility_std': 0.12180086970329285, 'reward_repeat_penalty_mean': 0.9554755091667175, 'reward_repeat_penalty_std': 0.03760902211070061, 'reward_near_duplicate_penalty_mean': 0.9656516313552856, 'reward_near_duplicate_penalty_std': 0.02569839358329773, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9893856048583984, 'reward_distinct_2_std': 0.025639133527874947, 'reward_judge_quality_mean': 0.4087499976158142, 'reward_judge_quality_std': 0.27445724606513977, 'reward_total_composite_mean': 0.3346109986305237, 'reward_total_composite_std': 0.28311118483543396, 'epoch': 0.02}

 15%|█▌        | 495/3300 [1:54:02<10:48:46, 13.88s/it]INFO 04-12 12:40:14 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 496/3300 [1:54:11<9:44:45, 12.51s/it] 2026-04-12 12:40:23,731 | INFO | train_grpo_train | metrics_logged mode=train step=496

                                                      
{'loss': 0.0155, 'grad_norm': 16.079675674438477, 'learning_rate': 8.5e-06, 'num_tokens': 982656.0, 'completions/mean_length': 77.125, 'completions/min_length': 69.0, 'completions/max_length': 82.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 77.125, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.9839693903923035, 'rewards/meter/std': 0.006644071079790592, '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/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.29999998211860657, 'rewards/judge_quality/std': 0.13093073666095734, 'rewards/repeat_penalty/mean': 0.5656132698059082, 'rewards/repeat_penalty/std': 0.28013280034065247, 'rewards/near_duplicate_penalty/mean': 0.8373399376869202, 'rewards/near_duplicate_penalty/std': 0.12468422204256058, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.7217047214508057, 'rewards/distinct_2/std': 0.18657812476158142, 'rewards/total_composite/mean': 0.2345302700996399, 'rewards/total_composite/std': 0.18441210687160492, 'reward': 0.2345302700996399, 'reward_std': 0.18441209197044373, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13972356915473938, 'sampling/sampling_logp_difference/max': 2.652090072631836, 'sampling/importance_sampling_ratio/min': 0.07050370424985886, 'sampling/importance_sampling_ratio/mean': 1.0060112476348877, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8917623087763786, 'clip_ratio/low_mean': 0.02820517448708415, 'clip_ratio/low_min': 0.02820517448708415, 'clip_ratio/high_mean': 0.07510748598724604, 'clip_ratio/high_max': 0.07510748598724604, 'clip_ratio/region_mean': 0.10331266047433019, 'reward_total_mean': 0.2345302700996399, 'reward_meter_mean': 0.9839693903923035, 'reward_meter_std': 0.006644071079790592, '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_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.5656132698059082, 'reward_repeat_penalty_std': 0.28013280034065247, 'reward_near_duplicate_penalty_mean': 0.8373399376869202, 'reward_near_duplicate_penalty_std': 0.12468422204256058, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.7217047214508057, 'reward_distinct_2_std': 0.18657812476158142, 'reward_judge_quality_mean': 0.29999998211860657, 'reward_judge_quality_std': 0.13093073666095734, 'reward_total_composite_mean': 0.2345302700996399, 'reward_total_composite_std': 0.18441210687160492, 'epoch': 0.02}

 15%|█▌        | 496/3300 [1:54:11<9:44:45, 12.51s/it]INFO 04-12 12:40:23 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 497/3300 [1:54:25<10:06:25, 12.98s/it]2026-04-12 12:40:37,803 | INFO | train_grpo_train | metrics_logged mode=train step=497

                                                       
{'loss': -0.2274, 'grad_norm': 1.3448582887649536, 'learning_rate': 8.496969696969697e-06, 'num_tokens': 986948.0, 'completions/mean_length': 404.5, 'completions/min_length': 315.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 389.14288330078125, 'completions/min_terminated_length': 315.0, 'completions/max_terminated_length': 483.0, 'rewards/meter/mean': 0.9185574054718018, 'rewards/meter/std': 0.20033182203769684, 'rewards/count_adherence/mean': 0.9107142686843872, 'rewards/count_adherence/std': 0.0739355981349945, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.13750000298023224, 'rewards/judge_quality/std': 0.0353553406894207, 'rewards/repeat_penalty/mean': 0.005796496756374836, 'rewards/repeat_penalty/std': 0.016394969075918198, 'rewards/near_duplicate_penalty/mean': 0.36620813608169556, 'rewards/near_duplicate_penalty/std': 0.2173338383436203, 'rewards/opening_diversity/mean': 0.4370805323123932, 'rewards/opening_diversity/std': 0.3685372769832611, 'rewards/distinct_2/mean': 0.009615384973585606, 'rewards/distinct_2/std': 0.02719641663134098, 'rewards/total_composite/mean': 0.11464738845825195, 'rewards/total_composite/std': 0.04099876806139946, 'reward': 0.11464738845825195, 'reward_std': 0.04099876433610916, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03906750679016113, 'sampling/sampling_logp_difference/max': 1.65081787109375, 'sampling/importance_sampling_ratio/min': 0.19189289212226868, 'sampling/importance_sampling_ratio/mean': 1.0020650625228882, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.29307733103632927, 'clip_ratio/low_mean': 0.0015873016091063619, 'clip_ratio/low_min': 0.0015873016091063619, 'clip_ratio/high_mean': 0.036969875916838646, 'clip_ratio/high_max': 0.036969875916838646, 'clip_ratio/region_mean': 0.03855717752594501, 'reward_total_mean': 0.11464738845825195, 'reward_meter_mean': 0.9185574054718018, 'reward_meter_std': 0.20033182203769684, 'reward_count_adherence_mean': 0.9107142686843872, 'reward_count_adherence_std': 0.0739355981349945, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.005796496756374836, 'reward_repeat_penalty_std': 0.016394969075918198, 'reward_near_duplicate_penalty_mean': 0.36620813608169556, 'reward_near_duplicate_penalty_std': 0.2173338383436203, 'reward_opening_diversity_mean': 0.4370805323123932, 'reward_opening_diversity_std': 0.3685372769832611, 'reward_distinct_2_mean': 0.009615384973585606, 'reward_distinct_2_std': 0.02719641663134098, 'reward_judge_quality_mean': 0.13750000298023224, 'reward_judge_quality_std': 0.0353553406894207, 'reward_total_composite_mean': 0.11464738845825195, 'reward_total_composite_std': 0.04099876806139946, 'epoch': 0.02}

 15%|█▌        | 497/3300 [1:54:25<10:06:25, 12.98s/it]INFO 04-12 12:40:38 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 498/3300 [1:54:39<10:16:01, 13.19s/it]2026-04-12 12:40:51,494 | INFO | train_grpo_train | metrics_logged mode=train step=498

                                                       
{'loss': -0.2485, 'grad_norm': 1.652599573135376, 'learning_rate': 8.493939393939394e-06, 'num_tokens': 989825.0, 'completions/mean_length': 280.625, 'completions/min_length': 157.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 203.5, 'completions/min_terminated_length': 157.0, 'completions/max_terminated_length': 250.0, 'rewards/meter/mean': 0.8432387113571167, 'rewards/meter/std': 0.3459304869174957, 'rewards/count_adherence/mean': 0.7777777910232544, 'rewards/count_adherence/std': 0.3894553780555725, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9609290361404419, 'rewards/lexical_plausibility/std': 0.11050929874181747, 'rewards/judge_quality/mean': 0.16250000894069672, 'rewards/judge_quality/std': 0.06408699601888657, 'rewards/repeat_penalty/mean': 0.2478998601436615, 'rewards/repeat_penalty/std': 0.43078354001045227, 'rewards/near_duplicate_penalty/mean': 0.6969571709632874, 'rewards/near_duplicate_penalty/std': 0.23057764768600464, 'rewards/opening_diversity/mean': 0.7395833134651184, 'rewards/opening_diversity/std': 0.2901747524738312, 'rewards/distinct_2/mean': 0.2668496370315552, 'rewards/distinct_2/std': 0.42589062452316284, 'rewards/total_composite/mean': 0.13889729976654053, 'rewards/total_composite/std': 0.08664173632860184, 'reward': 0.13889729976654053, 'reward_std': 0.08664173632860184, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05688926577568054, 'sampling/sampling_logp_difference/max': 1.4994902610778809, 'sampling/importance_sampling_ratio/min': 0.22324393689632416, 'sampling/importance_sampling_ratio/mean': 1.0079529285430908, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2630777284502983, 'clip_ratio/low_mean': 0.003980891779065132, 'clip_ratio/low_min': 0.003980891779065132, 'clip_ratio/high_mean': 0.03775472263805568, 'clip_ratio/high_max': 0.03775472263805568, 'clip_ratio/region_mean': 0.041735614417120814, 'reward_total_mean': 0.13889729976654053, 'reward_meter_mean': 0.8432387113571167, 'reward_meter_std': 0.3459304869174957, 'reward_count_adherence_mean': 0.7777777910232544, 'reward_count_adherence_std': 0.3894553780555725, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9609290361404419, 'reward_lexical_plausibility_std': 0.11050929874181747, 'reward_repeat_penalty_mean': 0.2478998601436615, 'reward_repeat_penalty_std': 0.43078354001045227, 'reward_near_duplicate_penalty_mean': 0.6969571709632874, 'reward_near_duplicate_penalty_std': 0.23057764768600464, 'reward_opening_diversity_mean': 0.7395833134651184, 'reward_opening_diversity_std': 0.2901747524738312, 'reward_distinct_2_mean': 0.2668496370315552, 'reward_distinct_2_std': 0.42589062452316284, 'reward_judge_quality_mean': 0.16250000894069672, 'reward_judge_quality_std': 0.06408699601888657, 'reward_total_composite_mean': 0.13889729976654053, 'reward_total_composite_std': 0.08664173632860184, 'epoch': 0.02}

 15%|█▌        | 498/3300 [1:54:39<10:16:01, 13.19s/it]INFO 04-12 12:40:51 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 499/3300 [1:54:48<9:13:31, 11.86s/it] 2026-04-12 12:41:00,228 | INFO | train_grpo_train | metrics_logged mode=train step=499

                                                      
{'loss': 0.0254, 'grad_norm': 12.778264045715332, 'learning_rate': 8.490909090909092e-06, 'num_tokens': 991688.0, 'completions/mean_length': 54.875, 'completions/min_length': 48.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.875, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.5963794589042664, 'rewards/meter/std': 0.4109093248844147, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.44624999165534973, 'rewards/judge_quality/std': 0.19668231904506683, 'rewards/repeat_penalty/mean': 0.9931745529174805, 'rewards/repeat_penalty/std': 0.010063727386295795, 'rewards/near_duplicate_penalty/mean': 0.9931745529174805, 'rewards/near_duplicate_penalty/std': 0.010063727386295795, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.22740943729877472, 'rewards/total_composite/std': 0.15244118869304657, 'reward': 0.22740943729877472, 'reward_std': 0.15244117379188538, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14866903424263, 'sampling/sampling_logp_difference/max': 1.9792890548706055, 'sampling/importance_sampling_ratio/min': 0.13816742599010468, 'sampling/importance_sampling_ratio/mean': 1.0106961727142334, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9755630791187286, 'clip_ratio/low_mean': 0.07379013486206532, 'clip_ratio/low_min': 0.07379013486206532, 'clip_ratio/high_mean': 0.06390305608510971, 'clip_ratio/high_max': 0.06390305608510971, 'clip_ratio/region_mean': 0.13769319094717503, 'reward_total_mean': 0.22740943729877472, 'reward_meter_mean': 0.5963794589042664, 'reward_meter_std': 0.4109093248844147, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9931745529174805, 'reward_repeat_penalty_std': 0.010063727386295795, 'reward_near_duplicate_penalty_mean': 0.9931745529174805, 'reward_near_duplicate_penalty_std': 0.010063727386295795, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.44624999165534973, 'reward_judge_quality_std': 0.19668231904506683, 'reward_total_composite_mean': 0.22740943729877472, 'reward_total_composite_std': 0.15244118869304657, 'epoch': 0.02}

 15%|█▌        | 499/3300 [1:54:48<9:13:31, 11.86s/it]INFO 04-12 12:41:00 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 500/3300 [1:54:57<8:32:18, 10.98s/it]2026-04-12 12:41:09,161 | INFO | train_grpo_train | metrics_logged mode=train step=500

                                                      
{'loss': 0.0793, 'grad_norm': 15.163819313049316, 'learning_rate': 8.487878787878789e-06, 'num_tokens': 993388.0, 'completions/mean_length': 54.5, 'completions/min_length': 43.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.5, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.5737890005111694, 'rewards/meter/std': 0.4020596146583557, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9678030014038086, 'rewards/lexical_plausibility/std': 0.06296154111623764, 'rewards/judge_quality/mean': 0.6050000190734863, 'rewards/judge_quality/std': 0.36637216806411743, 'rewards/repeat_penalty/mean': 0.7873538136482239, 'rewards/repeat_penalty/std': 0.2674378454685211, 'rewards/near_duplicate_penalty/mean': 0.9450277090072632, 'rewards/near_duplicate_penalty/std': 0.06491279602050781, 'rewards/opening_diversity/mean': 0.859375, 'rewards/opening_diversity/std': 0.2259652018547058, 'rewards/distinct_2/mean': 0.9507899284362793, 'rewards/distinct_2/std': 0.07793406397104263, 'rewards/total_composite/mean': 0.36765599250793457, 'rewards/total_composite/std': 0.3304150104522705, 'reward': 0.36765599250793457, 'reward_std': 0.3304150104522705, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1671167016029358, 'sampling/sampling_logp_difference/max': 1.783510684967041, 'sampling/importance_sampling_ratio/min': 0.16804715991020203, 'sampling/importance_sampling_ratio/mean': 1.024588704109192, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.223095528781414, 'clip_ratio/low_mean': 0.10117241274565458, 'clip_ratio/low_min': 0.10117241274565458, 'clip_ratio/high_mean': 0.06417830474674702, 'clip_ratio/high_max': 0.06417830474674702, 'clip_ratio/region_mean': 0.1653507174924016, 'reward_total_mean': 0.36765599250793457, 'reward_meter_mean': 0.5737890005111694, 'reward_meter_std': 0.4020596146583557, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9678030014038086, 'reward_lexical_plausibility_std': 0.06296154111623764, 'reward_repeat_penalty_mean': 0.7873538136482239, 'reward_repeat_penalty_std': 0.2674378454685211, 'reward_near_duplicate_penalty_mean': 0.9450277090072632, 'reward_near_duplicate_penalty_std': 0.06491279602050781, 'reward_opening_diversity_mean': 0.859375, 'reward_opening_diversity_std': 0.2259652018547058, 'reward_distinct_2_mean': 0.9507899284362793, 'reward_distinct_2_std': 0.07793406397104263, 'reward_judge_quality_mean': 0.6050000190734863, 'reward_judge_quality_std': 0.36637216806411743, 'reward_total_composite_mean': 0.36765599250793457, 'reward_total_composite_std': 0.3304150104522705, 'epoch': 0.02}

 15%|█▌        | 500/3300 [1:54:57<8:32:18, 10.98s/it]INFO 04-12 12:41:09 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


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

 15%|█▌        | 2/13 [00:11<01:02,  5.64s/it]

 23%|██▎       | 3/13 [00:18<01:02,  6.25s/it]

 31%|███       | 4/13 [00:29<01:13,  8.17s/it]

 38%|███▊      | 5/13 [00:40<01:12,  9.01s/it]

 46%|████▌     | 6/13 [00:52<01:09,  9.91s/it]

 54%|█████▍    | 7/13 [01:03<01:03, 10.53s/it]

 62%|██████▏   | 8/13 [01:18<00:59, 11.82s/it]

 69%|██████▉   | 9/13 [01:30<00:47, 11.90s/it]

 77%|███████▋  | 10/13 [01:42<00:35, 11.77s/it]

 85%|████████▍ | 11/13 [01:53<00:23, 11.75s/it]

 92%|█████████▏| 12/13 [02:04<00:11, 11.29s/it]

100%|██████████| 13/13 [02:15<00:00, 11.41s/it]2026-04-12 12:43:37,053 | INFO | train_grpo_train | metrics_logged mode=eval step=500

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 147.8471, 'eval_samples_per_second': 0.703, 'eval_steps_per_second': 0.088, 'eval_num_tokens': 993388.0, 'eval_completions/mean_length': 195.77884615384616, 'eval_completions/min_length': 48.0, 'eval_completions/max_length': 451.84615384615387, 'eval_completions/clipped_ratio': 0.07692307692307693, 'eval_completions/mean_terminated_length': 171.2225306584285, 'eval_completions/min_terminated_length': 48.0, 'eval_completions/max_terminated_length': 370.7692307692308, 'eval_rewards/meter/mean': 0.5651892286080581, 'eval_rewards/meter/std': 0.3941987959238199, 'eval_rewards/count_adherence/mean': 0.909946084022522, 'eval_rewards/count_adherence/std': 0.16018480793214762, 'eval_rewards/arabic_clean/mean': 0.875, 'eval_rewards/arabic_clean/std': 0.25542253255844116, 'eval_rewards/lexical_plausibility/mean': 0.9770224277789776, 'eval_rewards/lexical_plausibility/std': 0.057162257771079354, 'eval_rewards/judge_quality/mean': 0.37740384615384615, 'eval_rewards/judge_quality/std': 0.2514006575712791, 'eval_rewards/repeat_penalty/mean': 0.5576793345121237, 'eval_rewards/repeat_penalty/std': 0.42987576585549575, 'eval_rewards/near_duplicate_penalty/mean': 0.8533412401492779, 'eval_rewards/near_duplicate_penalty/std': 0.18654532386706427, 'eval_rewards/opening_diversity/mean': 0.883584031691918, 'eval_rewards/opening_diversity/std': 0.21873986950287452, 'eval_rewards/distinct_2/mean': 0.6092648460314825, 'eval_rewards/distinct_2/std': 0.4301047944105588, 'eval_rewards/total_composite/mean': 0.1936830941301126, 'eval_rewards/total_composite/std': 0.20124758694034356, 'eval_reward': 0.1936830941301126, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.05052991927816318, 'eval_sampling/sampling_logp_difference/max': 1.1012824498690093, 'eval_sampling/importance_sampling_ratio/min': 0.3416199019322029, 'eval_sampling/importance_sampling_ratio/mean': 1.0122336149215698, 'eval_sampling/importance_sampling_ratio/max': 1.4916566610336304, 'eval_entropy': 0.5890490962908819, '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': 0.1936830941301126, 'eval_reward_meter_mean': 0.5651892286080581, 'eval_reward_meter_std': 0.3941987959238199, 'eval_reward_count_adherence_mean': 0.909946084022522, 'eval_reward_count_adherence_std': 0.16018480793214762, 'eval_reward_arabic_clean_mean': 0.875, 'eval_reward_arabic_clean_std': 0.25542253255844116, 'eval_reward_lexical_plausibility_mean': 0.9770224277789776, 'eval_reward_lexical_plausibility_std': 0.057162257771079354, 'eval_reward_repeat_penalty_mean': 0.5576793345121237, 'eval_reward_repeat_penalty_std': 0.42987576585549575, 'eval_reward_near_duplicate_penalty_mean': 0.8533412401492779, 'eval_reward_near_duplicate_penalty_std': 0.18654532386706427, 'eval_reward_opening_diversity_mean': 0.883584031691918, 'eval_reward_opening_diversity_std': 0.21873986950287452, 'eval_reward_distinct_2_mean': 0.6092648460314825, 'eval_reward_distinct_2_std': 0.4301047944105588, 'eval_reward_judge_quality_mean': 0.37740384615384615, 'eval_reward_judge_quality_std': 0.2514006575712791, 'eval_reward_total_composite_mean': 0.1936830941301126, 'eval_reward_total_composite_std': 0.20124758694034356, 'epoch': 0.02}

 15%|█▌        | 500/3300 [1:57:25<8:32:18, 10.98s/it]

100%|██████████| 13/13 [02:15<00:00, 11.41s/it]

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

 15%|█▌        | 501/3300 [1:57:42<44:29:50, 57.23s/it]2026-04-12 12:43:54,343 | INFO | train_grpo_train | metrics_logged mode=train step=501

                                                       
{'loss': 0.0099, 'grad_norm': 2.6518707275390625, 'learning_rate': 8.484848484848486e-06, 'num_tokens': 998576.0, 'completions/mean_length': 387.5, 'completions/min_length': 319.0, 'completions/max_length': 465.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 387.5, 'completions/min_terminated_length': 319.0, 'completions/max_terminated_length': 465.0, 'rewards/meter/mean': 0.9701468348503113, 'rewards/meter/std': 0.02333131991326809, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.059608783572912216, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.97265625, 'rewards/lexical_plausibility/std': 0.07733980566263199, 'rewards/judge_quality/mean': 0.13750000298023224, 'rewards/judge_quality/std': 0.0353553406894207, 'rewards/repeat_penalty/mean': 0.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 0.6507840156555176, 'rewards/near_duplicate_penalty/std': 0.10026480257511139, 'rewards/opening_diversity/mean': 0.7787426710128784, 'rewards/opening_diversity/std': 0.18565993010997772, 'rewards/distinct_2/mean': 0.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.08965564519166946, 'rewards/total_composite/std': 0.06285939365625381, 'reward': 0.08965564519166946, 'reward_std': 0.06285939365625381, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.043643053621053696, 'sampling/sampling_logp_difference/max': 2.692099094390869, 'sampling/importance_sampling_ratio/min': 0.0677386000752449, 'sampling/importance_sampling_ratio/mean': 0.997916579246521, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2844799719750881, 'clip_ratio/low_mean': 0.012866653967648745, 'clip_ratio/low_min': 0.012866653967648745, 'clip_ratio/high_mean': 0.0384971653111279, 'clip_ratio/high_max': 0.0384971653111279, 'clip_ratio/region_mean': 0.051363819278776646, 'reward_total_mean': 0.08965564519166946, 'reward_meter_mean': 0.9701468348503113, 'reward_meter_std': 0.02333131991326809, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.059608783572912216, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.97265625, 'reward_lexical_plausibility_std': 0.07733980566263199, 'reward_repeat_penalty_mean': 0.0, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 0.6507840156555176, 'reward_near_duplicate_penalty_std': 0.10026480257511139, 'reward_opening_diversity_mean': 0.7787426710128784, 'reward_opening_diversity_std': 0.18565993010997772, 'reward_distinct_2_mean': 0.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.13750000298023224, 'reward_judge_quality_std': 0.0353553406894207, 'reward_total_composite_mean': 0.08965564519166946, 'reward_total_composite_std': 0.06285939365625381, 'epoch': 0.02}

 15%|█▌        | 501/3300 [1:57:42<44:29:50, 57.23s/it]INFO 04-12 12:43:54 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 502/3300 [1:57:56<34:22:35, 44.23s/it]2026-04-12 12:44:08,206 | INFO | train_grpo_train | metrics_logged mode=train step=502

                                                       
{'loss': -0.0317, 'grad_norm': 6.3878045082092285, 'learning_rate': 8.481818181818182e-06, 'num_tokens': 1000302.0, 'completions/mean_length': 99.75, 'completions/min_length': 35.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 40.85714340209961, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.3857800364494324, 'rewards/meter/std': 0.33045512437820435, '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/lexical_plausibility/mean': 0.9721478223800659, 'rewards/lexical_plausibility/std': 0.07877781242132187, 'rewards/judge_quality/mean': 0.4337499737739563, 'rewards/judge_quality/std': 0.23868314921855927, 'rewards/repeat_penalty/mean': 0.9405642747879028, 'rewards/repeat_penalty/std': 0.06924048066139221, 'rewards/near_duplicate_penalty/mean': 0.9689872860908508, 'rewards/near_duplicate_penalty/std': 0.042722415179014206, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9699149131774902, 'rewards/distinct_2/std': 0.041804179549217224, 'rewards/total_composite/mean': 0.21137171983718872, 'rewards/total_composite/std': 0.2944623529911041, 'reward': 0.21137171983718872, 'reward_std': 0.2944623529911041, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17319683730602264, 'sampling/sampling_logp_difference/max': 1.9168076515197754, 'sampling/importance_sampling_ratio/min': 0.14707574248313904, 'sampling/importance_sampling_ratio/mean': 1.0102839469909668, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7959209308028221, 'clip_ratio/low_mean': 0.07633984414860606, 'clip_ratio/low_min': 0.07633984414860606, 'clip_ratio/high_mean': 0.03692411910742521, 'clip_ratio/high_max': 0.03692411910742521, 'clip_ratio/region_mean': 0.11326396325603127, 'reward_total_mean': 0.21137171983718872, 'reward_meter_mean': 0.3857800364494324, 'reward_meter_std': 0.33045512437820435, '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_lexical_plausibility_mean': 0.9721478223800659, 'reward_lexical_plausibility_std': 0.07877781242132187, 'reward_repeat_penalty_mean': 0.9405642747879028, 'reward_repeat_penalty_std': 0.06924048066139221, 'reward_near_duplicate_penalty_mean': 0.9689872860908508, 'reward_near_duplicate_penalty_std': 0.042722415179014206, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9699149131774902, 'reward_distinct_2_std': 0.041804179549217224, 'reward_judge_quality_mean': 0.4337499737739563, 'reward_judge_quality_std': 0.23868314921855927, 'reward_total_composite_mean': 0.21137171983718872, 'reward_total_composite_std': 0.2944623529911041, 'epoch': 0.02}

 15%|█▌        | 502/3300 [1:57:56<34:22:35, 44.23s/it]INFO 04-12 12:44:08 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 503/3300 [1:58:05<26:15:41, 33.80s/it]2026-04-12 12:44:17,679 | INFO | train_grpo_train | metrics_logged mode=train step=503

                                                       
{'loss': 0.0662, 'grad_norm': 10.739781379699707, 'learning_rate': 8.478787878787879e-06, 'num_tokens': 1002102.0, 'completions/mean_length': 57.0, 'completions/min_length': 53.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.0, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.8123683333396912, 'rewards/meter/std': 0.3386837840080261, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.94140625, 'rewards/lexical_plausibility/std': 0.10881553590297699, 'rewards/judge_quality/mean': 0.4874999523162842, 'rewards/judge_quality/std': 0.1922609806060791, 'rewards/repeat_penalty/mean': 0.9695625901222229, 'rewards/repeat_penalty/std': 0.04207766056060791, 'rewards/near_duplicate_penalty/mean': 0.9759550094604492, 'rewards/near_duplicate_penalty/std': 0.027980675920844078, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9931318759918213, 'rewards/distinct_2/std': 0.01942601054906845, 'rewards/total_composite/mean': 0.41370487213134766, 'rewards/total_composite/std': 0.247430682182312, 'reward': 0.41370487213134766, 'reward_std': 0.247430682182312, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1602925807237625, 'sampling/sampling_logp_difference/max': 1.499053955078125, 'sampling/importance_sampling_ratio/min': 0.22334136068820953, 'sampling/importance_sampling_ratio/mean': 1.007265567779541, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3459157571196556, 'clip_ratio/low_mean': 0.04936777055263519, 'clip_ratio/low_min': 0.04936777055263519, 'clip_ratio/high_mean': 0.09928295202553272, 'clip_ratio/high_max': 0.09928295202553272, 'clip_ratio/region_mean': 0.14865072257816792, 'reward_total_mean': 0.41370487213134766, 'reward_meter_mean': 0.8123683333396912, 'reward_meter_std': 0.3386837840080261, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.94140625, 'reward_lexical_plausibility_std': 0.10881553590297699, 'reward_repeat_penalty_mean': 0.9695625901222229, 'reward_repeat_penalty_std': 0.04207766056060791, 'reward_near_duplicate_penalty_mean': 0.9759550094604492, 'reward_near_duplicate_penalty_std': 0.027980675920844078, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9931318759918213, 'reward_distinct_2_std': 0.01942601054906845, 'reward_judge_quality_mean': 0.4874999523162842, 'reward_judge_quality_std': 0.1922609806060791, 'reward_total_composite_mean': 0.41370487213134766, 'reward_total_composite_std': 0.247430682182312, 'epoch': 0.02}

 15%|█▌        | 503/3300 [1:58:05<26:15:41, 33.80s/it]INFO 04-12 12:44:17 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 504/3300 [1:58:14<20:30:36, 26.41s/it]2026-04-12 12:44:26,830 | INFO | train_grpo_train | metrics_logged mode=train step=504

                                                       
{'loss': 0.0069, 'grad_norm': 11.913764953613281, 'learning_rate': 8.475757575757576e-06, 'num_tokens': 1003762.0, 'completions/mean_length': 55.5, 'completions/min_length': 48.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.5, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.6863939166069031, 'rewards/meter/std': 0.40984946489334106, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.44999998807907104, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.7235284447669983, 'rewards/repeat_penalty/std': 0.2776219844818115, 'rewards/near_duplicate_penalty/mean': 0.8860005140304565, 'rewards/near_duplicate_penalty/std': 0.1023278534412384, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.8451676368713379, 'rewards/distinct_2/std': 0.19731010496616364, 'rewards/total_composite/mean': 0.3080506920814514, 'rewards/total_composite/std': 0.20203770697116852, 'reward': 0.3080506920814514, 'reward_std': 0.20203770697116852, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1357807219028473, 'sampling/sampling_logp_difference/max': 1.0764718055725098, 'sampling/importance_sampling_ratio/min': 0.3407957851886749, 'sampling/importance_sampling_ratio/mean': 1.0116885900497437, 'sampling/importance_sampling_ratio/max': 1.9978467226028442, 'entropy': 0.8812910094857216, 'clip_ratio/low_mean': 0.05885989125818014, 'clip_ratio/low_min': 0.05885989125818014, 'clip_ratio/high_mean': 0.08359778113663197, 'clip_ratio/high_max': 0.08359778113663197, 'clip_ratio/region_mean': 0.1424576723948121, 'reward_total_mean': 0.3080506920814514, 'reward_meter_mean': 0.6863939166069031, 'reward_meter_std': 0.40984946489334106, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7235284447669983, 'reward_repeat_penalty_std': 0.2776219844818115, 'reward_near_duplicate_penalty_mean': 0.8860005140304565, 'reward_near_duplicate_penalty_std': 0.1023278534412384, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.8451676368713379, 'reward_distinct_2_std': 0.19731010496616364, 'reward_judge_quality_mean': 0.44999998807907104, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.3080506920814514, 'reward_total_composite_std': 0.20203770697116852, 'epoch': 0.02}

 15%|█▌        | 504/3300 [1:58:14<20:30:36, 26.41s/it]INFO 04-12 12:44:27 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 505/3300 [1:58:28<17:37:30, 22.70s/it]2026-04-12 12:44:40,894 | INFO | train_grpo_train | metrics_logged mode=train step=505

                                                       
{'loss': -0.0951, 'grad_norm': 3.9400417804718018, 'learning_rate': 8.472727272727274e-06, 'num_tokens': 1007654.0, 'completions/mean_length': 343.5, 'completions/min_length': 253.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 319.4285888671875, 'completions/min_terminated_length': 253.0, 'completions/max_terminated_length': 402.0, 'rewards/meter/mean': 0.49392110109329224, 'rewards/meter/std': 0.4835190773010254, 'rewards/count_adherence/mean': 0.8181818723678589, 'rewards/count_adherence/std': 0.33313652873039246, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9263225793838501, 'rewards/lexical_plausibility/std': 0.12926451861858368, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.09258200973272324, 'rewards/repeat_penalty/mean': 0.17903828620910645, 'rewards/repeat_penalty/std': 0.3600805997848511, 'rewards/near_duplicate_penalty/mean': 0.5830976963043213, 'rewards/near_duplicate_penalty/std': 0.3056493401527405, 'rewards/opening_diversity/mean': 0.5512581467628479, 'rewards/opening_diversity/std': 0.3795725703239441, 'rewards/distinct_2/mean': 0.19717048108577728, 'rewards/distinct_2/std': 0.3628998100757599, 'rewards/total_composite/mean': 0.09154044836759567, 'rewards/total_composite/std': 0.11014432460069656, 'reward': 0.09154044836759567, 'reward_std': 0.11014432460069656, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06264153122901917, 'sampling/sampling_logp_difference/max': 2.0832653045654297, 'sampling/importance_sampling_ratio/min': 0.12452294677495956, 'sampling/importance_sampling_ratio/mean': 1.006455421447754, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.43054076842963696, 'clip_ratio/low_mean': 0.015363816171884537, 'clip_ratio/low_min': 0.015363816171884537, 'clip_ratio/high_mean': 0.03827769448980689, 'clip_ratio/high_max': 0.03827769448980689, 'clip_ratio/region_mean': 0.05364151066169143, 'reward_total_mean': 0.09154044836759567, 'reward_meter_mean': 0.49392110109329224, 'reward_meter_std': 0.4835190773010254, 'reward_count_adherence_mean': 0.8181818723678589, 'reward_count_adherence_std': 0.33313652873039246, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9263225793838501, 'reward_lexical_plausibility_std': 0.12926451861858368, 'reward_repeat_penalty_mean': 0.17903828620910645, 'reward_repeat_penalty_std': 0.3600805997848511, 'reward_near_duplicate_penalty_mean': 0.5830976963043213, 'reward_near_duplicate_penalty_std': 0.3056493401527405, 'reward_opening_diversity_mean': 0.5512581467628479, 'reward_opening_diversity_std': 0.3795725703239441, 'reward_distinct_2_mean': 0.19717048108577728, 'reward_distinct_2_std': 0.3628998100757599, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.09258200973272324, 'reward_total_composite_mean': 0.09154044836759567, 'reward_total_composite_std': 0.11014432460069656, 'epoch': 0.02}

 15%|█▌        | 505/3300 [1:58:28<17:37:30, 22.70s/it]INFO 04-12 12:44:41 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 506/3300 [1:58:36<14:04:44, 18.14s/it]2026-04-12 12:44:48,379 | INFO | train_grpo_train | metrics_logged mode=train step=506

                                                       
{'loss': 0.0633, 'grad_norm': 18.310579299926758, 'learning_rate': 8.46969696969697e-06, 'num_tokens': 1009037.0, 'completions/mean_length': 23.875, 'completions/min_length': 21.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.875, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.43831267952919006, 'rewards/meter/std': 0.28032413125038147, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.7387499809265137, 'rewards/judge_quality/std': 0.3636496663093567, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.31349655985832214, 'rewards/total_composite/std': 0.19792938232421875, 'reward': 0.31349655985832214, 'reward_std': 0.19792936742305756, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12091462314128876, 'sampling/sampling_logp_difference/max': 0.961705207824707, 'sampling/importance_sampling_ratio/min': 0.38224053382873535, 'sampling/importance_sampling_ratio/mean': 1.0261863470077515, 'sampling/importance_sampling_ratio/max': 1.9883272647857666, 'entropy': 0.9679062403738499, 'clip_ratio/low_mean': 0.06016457639634609, 'clip_ratio/low_min': 0.06016457639634609, 'clip_ratio/high_mean': 0.06392045505344868, 'clip_ratio/high_max': 0.06392045505344868, 'clip_ratio/region_mean': 0.12408503144979477, 'reward_total_mean': 0.31349655985832214, 'reward_meter_mean': 0.43831267952919006, 'reward_meter_std': 0.28032413125038147, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7387499809265137, 'reward_judge_quality_std': 0.3636496663093567, 'reward_total_composite_mean': 0.31349655985832214, 'reward_total_composite_std': 0.19792938232421875, 'epoch': 0.02}

 15%|█▌        | 506/3300 [1:58:36<14:04:44, 18.14s/it]INFO 04-12 12:44:48 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 507/3300 [1:58:49<12:53:24, 16.61s/it]2026-04-12 12:45:01,440 | INFO | train_grpo_train | metrics_logged mode=train step=507

                                                       
{'loss': -0.0479, 'grad_norm': 2.836268663406372, 'learning_rate': 8.466666666666668e-06, 'num_tokens': 1010449.0, 'completions/mean_length': 157.5, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 39.333335876464844, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.29638734459877014, 'rewards/meter/std': 0.3345950245857239, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.3720119297504425, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9288312792778015, 'rewards/lexical_plausibility/std': 0.1276407390832901, 'rewards/judge_quality/mean': 0.4950000047683716, 'rewards/judge_quality/std': 0.38340580463409424, 'rewards/repeat_penalty/mean': 0.9603478908538818, 'rewards/repeat_penalty/std': 0.08630295842885971, 'rewards/near_duplicate_penalty/mean': 0.9915978908538818, 'rewards/near_duplicate_penalty/std': 0.01535733137279749, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.18346711993217468, 'rewards/total_composite/std': 0.2076093703508377, 'reward': 0.18346711993217468, 'reward_std': 0.2076093703508377, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16101914644241333, 'sampling/sampling_logp_difference/max': 1.877507209777832, 'sampling/importance_sampling_ratio/min': 0.15297095477581024, 'sampling/importance_sampling_ratio/mean': 0.991775631904602, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0024839416146278, 'clip_ratio/low_mean': 0.053816527128219604, 'clip_ratio/low_min': 0.053816527128219604, 'clip_ratio/high_mean': 0.06941848620772362, 'clip_ratio/high_max': 0.06941848620772362, 'clip_ratio/region_mean': 0.12323501333594322, 'reward_total_mean': 0.18346711993217468, 'reward_meter_mean': 0.29638734459877014, 'reward_meter_std': 0.3345950245857239, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.3720119297504425, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9288312792778015, 'reward_lexical_plausibility_std': 0.1276407390832901, 'reward_repeat_penalty_mean': 0.9603478908538818, 'reward_repeat_penalty_std': 0.08630295842885971, 'reward_near_duplicate_penalty_mean': 0.9915978908538818, 'reward_near_duplicate_penalty_std': 0.01535733137279749, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4950000047683716, 'reward_judge_quality_std': 0.38340580463409424, 'reward_total_composite_mean': 0.18346711993217468, 'reward_total_composite_std': 0.2076093703508377, 'epoch': 0.02}

 15%|█▌        | 507/3300 [1:58:49<12:53:24, 16.61s/it]INFO 04-12 12:45:01 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 508/3300 [1:59:03<12:17:38, 15.85s/it]2026-04-12 12:45:15,509 | INFO | train_grpo_train | metrics_logged mode=train step=508

                                                       
{'loss': -0.166, 'grad_norm': 1.8281652927398682, 'learning_rate': 8.463636363636364e-06, 'num_tokens': 1014431.0, 'completions/mean_length': 350.75, 'completions/min_length': 285.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 327.71429443359375, 'completions/min_terminated_length': 285.0, 'completions/max_terminated_length': 347.0, 'rewards/meter/mean': 0.9088338017463684, 'rewards/meter/std': 0.23093324899673462, 'rewards/count_adherence/mean': 0.8229166865348816, 'rewards/count_adherence/std': 0.26888033747673035, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.989179790019989, 'rewards/lexical_plausibility/std': 0.030604196712374687, 'rewards/judge_quality/mean': 0.13125000894069672, 'rewards/judge_quality/std': 0.0530330128967762, 'rewards/repeat_penalty/mean': 0.0658489540219307, 'rewards/repeat_penalty/std': 0.18624897301197052, 'rewards/near_duplicate_penalty/mean': 0.4798297882080078, 'rewards/near_duplicate_penalty/std': 0.2374846488237381, 'rewards/opening_diversity/mean': 0.541671097278595, 'rewards/opening_diversity/std': 0.27736055850982666, 'rewards/distinct_2/mean': 0.11935897171497345, 'rewards/distinct_2/std': 0.33759817481040955, 'rewards/total_composite/mean': 0.09086012095212936, 'rewards/total_composite/std': 0.06445092707872391, 'reward': 0.09086012095212936, 'reward_std': 0.0644509345293045, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03532516583800316, 'sampling/sampling_logp_difference/max': 1.8691463470458984, 'sampling/importance_sampling_ratio/min': 0.15425528585910797, 'sampling/importance_sampling_ratio/mean': 1.0060081481933594, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1880291998386383, 'clip_ratio/low_mean': 0.007126605371013284, 'clip_ratio/low_min': 0.007126605371013284, 'clip_ratio/high_mean': 0.015821539680473506, 'clip_ratio/high_max': 0.015821539680473506, 'clip_ratio/region_mean': 0.02294814505148679, 'reward_total_mean': 0.09086012095212936, 'reward_meter_mean': 0.9088338017463684, 'reward_meter_std': 0.23093324899673462, 'reward_count_adherence_mean': 0.8229166865348816, 'reward_count_adherence_std': 0.26888033747673035, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.989179790019989, 'reward_lexical_plausibility_std': 0.030604196712374687, 'reward_repeat_penalty_mean': 0.0658489540219307, 'reward_repeat_penalty_std': 0.18624897301197052, 'reward_near_duplicate_penalty_mean': 0.4798297882080078, 'reward_near_duplicate_penalty_std': 0.2374846488237381, 'reward_opening_diversity_mean': 0.541671097278595, 'reward_opening_diversity_std': 0.27736055850982666, 'reward_distinct_2_mean': 0.11935897171497345, 'reward_distinct_2_std': 0.33759817481040955, 'reward_judge_quality_mean': 0.13125000894069672, 'reward_judge_quality_std': 0.0530330128967762, 'reward_total_composite_mean': 0.09086012095212936, 'reward_total_composite_std': 0.06445092707872391, 'epoch': 0.02}

 15%|█▌        | 508/3300 [1:59:03<12:17:38, 15.85s/it]INFO 04-12 12:45:15 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 509/3300 [1:59:17<11:49:59, 15.26s/it]2026-04-12 12:45:29,400 | INFO | train_grpo_train | metrics_logged mode=train step=509

                                                       
{'loss': -0.0936, 'grad_norm': 5.034311294555664, 'learning_rate': 8.460606060606061e-06, 'num_tokens': 1015936.0, 'completions/mean_length': 109.125, 'completions/min_length': 45.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 51.57143020629883, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.945076048374176, 'rewards/meter/std': 0.04943845793604851, '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/lexical_plausibility/mean': 0.9866071343421936, 'rewards/lexical_plausibility/std': 0.03788072615861893, 'rewards/judge_quality/mean': 0.3500000238418579, 'rewards/judge_quality/std': 0.20701967179775238, 'rewards/repeat_penalty/mean': 0.761979341506958, 'rewards/repeat_penalty/std': 0.33365947008132935, 'rewards/near_duplicate_penalty/mean': 0.9482151865959167, 'rewards/near_duplicate_penalty/std': 0.04744406417012215, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.829497218132019, 'rewards/distinct_2/std': 0.3429631292819977, 'rewards/total_composite/mean': 0.32543379068374634, 'rewards/total_composite/std': 0.19852536916732788, 'reward': 0.32543379068374634, 'reward_std': 0.19852536916732788, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16109664738178253, 'sampling/sampling_logp_difference/max': 2.4872822761535645, 'sampling/importance_sampling_ratio/min': 0.08313559740781784, 'sampling/importance_sampling_ratio/mean': 0.9837594032287598, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8953721970319748, 'clip_ratio/low_mean': 0.045713142259046435, 'clip_ratio/low_min': 0.045713142259046435, 'clip_ratio/high_mean': 0.11555221676826477, 'clip_ratio/high_max': 0.11555221676826477, 'clip_ratio/region_mean': 0.1612653590273112, 'reward_total_mean': 0.32543379068374634, 'reward_meter_mean': 0.945076048374176, 'reward_meter_std': 0.04943845793604851, '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_lexical_plausibility_mean': 0.9866071343421936, 'reward_lexical_plausibility_std': 0.03788072615861893, 'reward_repeat_penalty_mean': 0.761979341506958, 'reward_repeat_penalty_std': 0.33365947008132935, 'reward_near_duplicate_penalty_mean': 0.9482151865959167, 'reward_near_duplicate_penalty_std': 0.04744406417012215, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.829497218132019, 'reward_distinct_2_std': 0.3429631292819977, 'reward_judge_quality_mean': 0.3500000238418579, 'reward_judge_quality_std': 0.20701967179775238, 'reward_total_composite_mean': 0.32543379068374634, 'reward_total_composite_std': 0.19852536916732788, 'epoch': 0.02}

 15%|█▌        | 509/3300 [1:59:17<11:49:59, 15.26s/it]INFO 04-12 12:45:29 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 510/3300 [1:59:32<11:53:40, 15.35s/it]2026-04-12 12:45:44,966 | INFO | train_grpo_train | metrics_logged mode=train step=510

                                                       
{'loss': 0.0131, 'grad_norm': 6.778062343597412, 'learning_rate': 8.457575757575758e-06, 'num_tokens': 1018883.0, 'completions/mean_length': 173.375, 'completions/min_length': 116.0, 'completions/max_length': 216.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 173.375, 'completions/min_terminated_length': 116.0, 'completions/max_terminated_length': 216.0, 'rewards/meter/mean': 0.7390681505203247, 'rewards/meter/std': 0.3173445165157318, 'rewards/count_adherence/mean': 0.9464285969734192, 'rewards/count_adherence/std': 0.10628911107778549, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9866071343421936, 'rewards/lexical_plausibility/std': 0.03788072615861893, 'rewards/judge_quality/mean': 0.23750001192092896, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.2772557735443115, 'rewards/repeat_penalty/std': 0.29759883880615234, 'rewards/near_duplicate_penalty/mean': 0.7736907005310059, 'rewards/near_duplicate_penalty/std': 0.11297912895679474, 'rewards/opening_diversity/mean': 0.7393543720245361, 'rewards/opening_diversity/std': 0.2959821820259094, 'rewards/distinct_2/mean': 0.4010602831840515, 'rewards/distinct_2/std': 0.2960413992404938, 'rewards/total_composite/mean': 0.16854535043239594, 'rewards/total_composite/std': 0.10978459566831589, 'reward': 0.16854535043239594, 'reward_std': 0.10978458821773529, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0983595922589302, 'sampling/sampling_logp_difference/max': 3.287526845932007, 'sampling/importance_sampling_ratio/min': 0.03734609857201576, 'sampling/importance_sampling_ratio/mean': 1.0133074522018433, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6525220572948456, 'clip_ratio/low_mean': 0.05915332958102226, 'clip_ratio/low_min': 0.05915332958102226, 'clip_ratio/high_mean': 0.023187135346233845, 'clip_ratio/high_max': 0.023187135346233845, 'clip_ratio/region_mean': 0.08234046492725611, 'reward_total_mean': 0.16854535043239594, 'reward_meter_mean': 0.7390681505203247, 'reward_meter_std': 0.3173445165157318, 'reward_count_adherence_mean': 0.9464285969734192, 'reward_count_adherence_std': 0.10628911107778549, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9866071343421936, 'reward_lexical_plausibility_std': 0.03788072615861893, 'reward_repeat_penalty_mean': 0.2772557735443115, 'reward_repeat_penalty_std': 0.29759883880615234, 'reward_near_duplicate_penalty_mean': 0.7736907005310059, 'reward_near_duplicate_penalty_std': 0.11297912895679474, 'reward_opening_diversity_mean': 0.7393543720245361, 'reward_opening_diversity_std': 0.2959821820259094, 'reward_distinct_2_mean': 0.4010602831840515, 'reward_distinct_2_std': 0.2960413992404938, 'reward_judge_quality_mean': 0.23750001192092896, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.16854535043239594, 'reward_total_composite_std': 0.10978459566831589, 'epoch': 0.02}

 15%|█▌        | 510/3300 [1:59:32<11:53:40, 15.35s/it]INFO 04-12 12:45:45 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 511/3300 [1:59:45<11:08:46, 14.39s/it]2026-04-12 12:45:57,094 | INFO | train_grpo_train | metrics_logged mode=train step=511

                                                       
{'loss': 0.0254, 'grad_norm': 5.904370307922363, 'learning_rate': 8.454545454545455e-06, 'num_tokens': 1022390.0, 'completions/mean_length': 239.375, 'completions/min_length': 215.0, 'completions/max_length': 314.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 239.375, 'completions/min_terminated_length': 215.0, 'completions/max_terminated_length': 314.0, 'rewards/meter/mean': 0.9285152554512024, 'rewards/meter/std': 0.12181481719017029, 'rewards/count_adherence/mean': 0.9305555820465088, 'rewards/count_adherence/std': 0.05750546231865883, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.21249999105930328, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.16502662003040314, 'rewards/repeat_penalty/std': 0.2941958010196686, 'rewards/near_duplicate_penalty/mean': 0.7388197779655457, 'rewards/near_duplicate_penalty/std': 0.1587085872888565, 'rewards/opening_diversity/mean': 0.8895833492279053, 'rewards/opening_diversity/std': 0.09393831342458725, 'rewards/distinct_2/mean': 0.18130643665790558, 'rewards/distinct_2/std': 0.3143492043018341, 'rewards/total_composite/mean': 0.180444598197937, 'rewards/total_composite/std': 0.10672377049922943, 'reward': 0.180444598197937, 'reward_std': 0.10672377049922943, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07582195103168488, 'sampling/sampling_logp_difference/max': 2.2196154594421387, 'sampling/importance_sampling_ratio/min': 0.10865087807178497, 'sampling/importance_sampling_ratio/mean': 1.0033222436904907, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4280640445649624, 'clip_ratio/low_mean': 0.03521737502887845, 'clip_ratio/low_min': 0.03521737502887845, 'clip_ratio/high_mean': 0.031118144281208515, 'clip_ratio/high_max': 0.031118144281208515, 'clip_ratio/region_mean': 0.06633551931008697, 'reward_total_mean': 0.180444598197937, 'reward_meter_mean': 0.9285152554512024, 'reward_meter_std': 0.12181481719017029, 'reward_count_adherence_mean': 0.9305555820465088, 'reward_count_adherence_std': 0.05750546231865883, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.16502662003040314, 'reward_repeat_penalty_std': 0.2941958010196686, 'reward_near_duplicate_penalty_mean': 0.7388197779655457, 'reward_near_duplicate_penalty_std': 0.1587085872888565, 'reward_opening_diversity_mean': 0.8895833492279053, 'reward_opening_diversity_std': 0.09393831342458725, 'reward_distinct_2_mean': 0.18130643665790558, 'reward_distinct_2_std': 0.3143492043018341, 'reward_judge_quality_mean': 0.21249999105930328, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.180444598197937, 'reward_total_composite_std': 0.10672377049922943, 'epoch': 0.02}

 15%|█▌        | 511/3300 [1:59:45<11:08:46, 14.39s/it]INFO 04-12 12:45:57 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 512/3300 [1:59:54<10:03:10, 12.98s/it]2026-04-12 12:46:06,797 | INFO | train_grpo_train | metrics_logged mode=train step=512

                                                       
{'loss': 0.0524, 'grad_norm': 8.661601066589355, 'learning_rate': 8.451515151515151e-06, 'num_tokens': 1024517.0, 'completions/mean_length': 102.875, 'completions/min_length': 84.0, 'completions/max_length': 137.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 102.875, 'completions/min_terminated_length': 84.0, 'completions/max_terminated_length': 137.0, 'rewards/meter/mean': 0.8079421520233154, 'rewards/meter/std': 0.2958540916442871, '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/lexical_plausibility/mean': 0.9397321343421936, 'rewards/lexical_plausibility/std': 0.13258251547813416, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.3515321910381317, 'rewards/repeat_penalty/std': 0.2766278088092804, 'rewards/near_duplicate_penalty/mean': 0.8433783054351807, 'rewards/near_duplicate_penalty/std': 0.08072208613157272, 'rewards/opening_diversity/mean': 0.8504464626312256, 'rewards/opening_diversity/std': 0.20014743506908417, 'rewards/distinct_2/mean': 0.5110594034194946, 'rewards/distinct_2/std': 0.2516166865825653, 'rewards/total_composite/mean': 0.1917930245399475, 'rewards/total_composite/std': 0.12354664504528046, 'reward': 0.1917930245399475, 'reward_std': 0.12354663759469986, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09449737519025803, 'sampling/sampling_logp_difference/max': 1.7361695766448975, 'sampling/importance_sampling_ratio/min': 0.17619401216506958, 'sampling/importance_sampling_ratio/mean': 1.0089701414108276, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6085465885698795, 'clip_ratio/low_mean': 0.05178106855601072, 'clip_ratio/low_min': 0.05178106855601072, 'clip_ratio/high_mean': 0.0491557726636529, 'clip_ratio/high_max': 0.0491557726636529, 'clip_ratio/region_mean': 0.10093684121966362, 'reward_total_mean': 0.1917930245399475, 'reward_meter_mean': 0.8079421520233154, 'reward_meter_std': 0.2958540916442871, '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_lexical_plausibility_mean': 0.9397321343421936, 'reward_lexical_plausibility_std': 0.13258251547813416, 'reward_repeat_penalty_mean': 0.3515321910381317, 'reward_repeat_penalty_std': 0.2766278088092804, 'reward_near_duplicate_penalty_mean': 0.8433783054351807, 'reward_near_duplicate_penalty_std': 0.08072208613157272, 'reward_opening_diversity_mean': 0.8504464626312256, 'reward_opening_diversity_std': 0.20014743506908417, 'reward_distinct_2_mean': 0.5110594034194946, 'reward_distinct_2_std': 0.2516166865825653, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.1917930245399475, 'reward_total_composite_std': 0.12354664504528046, 'epoch': 0.02}

 16%|█▌        | 512/3300 [1:59:54<10:03:10, 12.98s/it]INFO 04-12 12:46:07 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 513/3300 [2:00:03<9:05:25, 11.74s/it] 2026-04-12 12:46:15,643 | INFO | train_grpo_train | metrics_logged mode=train step=513

                                                      
{'loss': 0.0136, 'grad_norm': 16.81121253967285, 'learning_rate': 8.44848484848485e-06, 'num_tokens': 1026221.0, 'completions/mean_length': 45.0, 'completions/min_length': 42.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.0, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.4247587323188782, 'rewards/meter/std': 0.4042350649833679, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5924999713897705, 'rewards/judge_quality/std': 0.3535837233066559, 'rewards/repeat_penalty/mean': 0.9381329417228699, 'rewards/repeat_penalty/std': 0.11472293734550476, 'rewards/near_duplicate_penalty/mean': 0.9840990900993347, 'rewards/near_duplicate_penalty/std': 0.014699822291731834, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9811372756958008, 'rewards/distinct_2/std': 0.035051409155130386, 'rewards/total_composite/mean': 0.18739627301692963, 'rewards/total_composite/std': 0.2175557166337967, 'reward': 0.18739627301692963, 'reward_std': 0.21755573153495789, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15761037170886993, 'sampling/sampling_logp_difference/max': 1.3418798446655273, 'sampling/importance_sampling_ratio/min': 0.26135390996932983, 'sampling/importance_sampling_ratio/mean': 1.0226404666900635, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0555620938539505, 'clip_ratio/low_mean': 0.12172575201839209, 'clip_ratio/low_min': 0.12172575201839209, 'clip_ratio/high_mean': 0.03297101520001888, 'clip_ratio/high_max': 0.03297101520001888, 'clip_ratio/region_mean': 0.15469676721841097, 'reward_total_mean': 0.18739627301692963, 'reward_meter_mean': 0.4247587323188782, 'reward_meter_std': 0.4042350649833679, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9381329417228699, 'reward_repeat_penalty_std': 0.11472293734550476, 'reward_near_duplicate_penalty_mean': 0.9840990900993347, 'reward_near_duplicate_penalty_std': 0.014699822291731834, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9811372756958008, 'reward_distinct_2_std': 0.035051409155130386, 'reward_judge_quality_mean': 0.5924999713897705, 'reward_judge_quality_std': 0.3535837233066559, 'reward_total_composite_mean': 0.18739627301692963, 'reward_total_composite_std': 0.2175557166337967, 'epoch': 0.02}

 16%|█▌        | 513/3300 [2:00:03<9:05:25, 11.74s/it]INFO 04-12 12:46:15 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 514/3300 [2:00:17<9:37:00, 12.43s/it]2026-04-12 12:46:29,691 | INFO | train_grpo_train | metrics_logged mode=train step=514

                                                      
{'loss': -0.0577, 'grad_norm': 5.431741714477539, 'learning_rate': 8.445454545454547e-06, 'num_tokens': 1028079.0, 'completions/mean_length': 110.25, 'completions/min_length': 45.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 52.85714340209961, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.513384997844696, 'rewards/meter/std': 0.40638771653175354, '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/lexical_plausibility/mean': 0.949152946472168, 'rewards/lexical_plausibility/std': 0.13294920325279236, 'rewards/judge_quality/mean': 0.5762499570846558, 'rewards/judge_quality/std': 0.32845690846443176, 'rewards/repeat_penalty/mean': 0.9661033153533936, 'rewards/repeat_penalty/std': 0.0875559076666832, 'rewards/near_duplicate_penalty/mean': 0.9973533153533936, 'rewards/near_duplicate_penalty/std': 0.006525635253638029, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3251316547393799, 'rewards/total_composite/std': 0.32359376549720764, 'reward': 0.3251316547393799, 'reward_std': 0.32359376549720764, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19153591990470886, 'sampling/sampling_logp_difference/max': 2.0924038887023926, 'sampling/importance_sampling_ratio/min': 0.12339016795158386, 'sampling/importance_sampling_ratio/mean': 1.0208725929260254, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3402554169297218, 'clip_ratio/low_mean': 0.06283831968903542, 'clip_ratio/low_min': 0.06283831968903542, 'clip_ratio/high_mean': 0.08408079575747252, 'clip_ratio/high_max': 0.08408079575747252, 'clip_ratio/region_mean': 0.14691911544650793, 'reward_total_mean': 0.3251316547393799, 'reward_meter_mean': 0.513384997844696, 'reward_meter_std': 0.40638771653175354, '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_lexical_plausibility_mean': 0.949152946472168, 'reward_lexical_plausibility_std': 0.13294920325279236, 'reward_repeat_penalty_mean': 0.9661033153533936, 'reward_repeat_penalty_std': 0.0875559076666832, 'reward_near_duplicate_penalty_mean': 0.9973533153533936, 'reward_near_duplicate_penalty_std': 0.006525635253638029, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5762499570846558, 'reward_judge_quality_std': 0.32845690846443176, 'reward_total_composite_mean': 0.3251316547393799, 'reward_total_composite_std': 0.32359376549720764, 'epoch': 0.02}

 16%|█▌        | 514/3300 [2:00:17<9:37:00, 12.43s/it]INFO 04-12 12:46:29 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 515/3300 [2:00:31<9:51:45, 12.75s/it]2026-04-12 12:46:43,166 | INFO | train_grpo_train | metrics_logged mode=train step=515

                                                      
{'loss': -0.0291, 'grad_norm': 8.370580673217773, 'learning_rate': 8.442424242424243e-06, 'num_tokens': 1029549.0, 'completions/mean_length': 100.75, 'completions/min_length': 37.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 42.0, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.36182424426078796, 'rewards/meter/std': 0.39704596996307373, '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/lexical_plausibility/mean': 0.9768381714820862, 'rewards/lexical_plausibility/std': 0.04581486061215401, 'rewards/judge_quality/mean': 0.6262500286102295, 'rewards/judge_quality/std': 0.3452095687389374, 'rewards/repeat_penalty/mean': 0.977394163608551, 'rewards/repeat_penalty/std': 0.027574090287089348, 'rewards/near_duplicate_penalty/mean': 0.977394163608551, 'rewards/near_duplicate_penalty/std': 0.027574090287089348, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.22116199135780334, 'rewards/total_composite/std': 0.2753816545009613, 'reward': 0.22116199135780334, 'reward_std': 0.2753816545009613, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14220990240573883, 'sampling/sampling_logp_difference/max': 1.3544740676879883, 'sampling/importance_sampling_ratio/min': 0.2580829858779907, 'sampling/importance_sampling_ratio/mean': 1.0478286743164062, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7890208065509796, 'clip_ratio/low_mean': 0.046606277115643024, 'clip_ratio/low_min': 0.046606277115643024, 'clip_ratio/high_mean': 0.054215616546571255, 'clip_ratio/high_max': 0.054215616546571255, 'clip_ratio/region_mean': 0.10082189366221428, 'reward_total_mean': 0.22116199135780334, 'reward_meter_mean': 0.36182424426078796, 'reward_meter_std': 0.39704596996307373, '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_lexical_plausibility_mean': 0.9768381714820862, 'reward_lexical_plausibility_std': 0.04581486061215401, 'reward_repeat_penalty_mean': 0.977394163608551, 'reward_repeat_penalty_std': 0.027574090287089348, 'reward_near_duplicate_penalty_mean': 0.977394163608551, 'reward_near_duplicate_penalty_std': 0.027574090287089348, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6262500286102295, 'reward_judge_quality_std': 0.3452095687389374, 'reward_total_composite_mean': 0.22116199135780334, 'reward_total_composite_std': 0.2753816545009613, 'epoch': 0.02}

 16%|█▌        | 515/3300 [2:00:31<9:51:45, 12.75s/it]INFO 04-12 12:46:43 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 516/3300 [2:00:45<10:12:00, 13.19s/it]2026-04-12 12:46:57,383 | INFO | train_grpo_train | metrics_logged mode=train step=516

                                                       
{'loss': -0.0737, 'grad_norm': 5.375249862670898, 'learning_rate': 8.43939393939394e-06, 'num_tokens': 1031237.0, 'completions/mean_length': 109.0, 'completions/min_length': 44.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 51.42857360839844, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.34669095277786255, 'rewards/meter/std': 0.24385115504264832, '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/lexical_plausibility/mean': 0.9390032291412354, 'rewards/lexical_plausibility/std': 0.09921156615018845, 'rewards/judge_quality/mean': 0.5174999833106995, 'rewards/judge_quality/std': 0.2841905951499939, 'rewards/repeat_penalty/mean': 0.9837115406990051, 'rewards/repeat_penalty/std': 0.037624627351760864, 'rewards/near_duplicate_penalty/mean': 0.9935320615768433, 'rewards/near_duplicate_penalty/std': 0.011036964133381844, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9898785352706909, 'rewards/distinct_2/std': 0.028627805411815643, 'rewards/total_composite/mean': 0.15397033095359802, 'rewards/total_composite/std': 0.12483875453472137, 'reward': 0.15397033095359802, 'reward_std': 0.12483874708414078, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17727911472320557, 'sampling/sampling_logp_difference/max': 1.3981847763061523, 'sampling/importance_sampling_ratio/min': 0.24704501032829285, 'sampling/importance_sampling_ratio/mean': 1.0220539569854736, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.154013380408287, 'clip_ratio/low_mean': 0.061760082840919495, 'clip_ratio/low_min': 0.061760082840919495, 'clip_ratio/high_mean': 0.08354979753494263, 'clip_ratio/high_max': 0.08354979753494263, 'clip_ratio/region_mean': 0.14530988037586212, 'reward_total_mean': 0.15397033095359802, 'reward_meter_mean': 0.34669095277786255, 'reward_meter_std': 0.24385115504264832, '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_lexical_plausibility_mean': 0.9390032291412354, 'reward_lexical_plausibility_std': 0.09921156615018845, 'reward_repeat_penalty_mean': 0.9837115406990051, 'reward_repeat_penalty_std': 0.037624627351760864, 'reward_near_duplicate_penalty_mean': 0.9935320615768433, 'reward_near_duplicate_penalty_std': 0.011036964133381844, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9898785352706909, 'reward_distinct_2_std': 0.028627805411815643, 'reward_judge_quality_mean': 0.5174999833106995, 'reward_judge_quality_std': 0.2841905951499939, 'reward_total_composite_mean': 0.15397033095359802, 'reward_total_composite_std': 0.12483875453472137, 'epoch': 0.02}

 16%|█▌        | 516/3300 [2:00:45<10:12:00, 13.19s/it]INFO 04-12 12:46:57 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 517/3300 [2:01:02<11:03:04, 14.30s/it]2026-04-12 12:47:14,264 | INFO | train_grpo_train | metrics_logged mode=train step=517

                                                       
{'loss': -0.0667, 'grad_norm': 3.699153184890747, 'learning_rate': 8.436363636363637e-06, 'num_tokens': 1034315.0, 'completions/mean_length': 217.75, 'completions/min_length': 133.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 175.71429443359375, 'completions/min_terminated_length': 133.0, 'completions/max_terminated_length': 245.0, 'rewards/meter/mean': 0.801067590713501, 'rewards/meter/std': 0.18090085685253143, 'rewards/count_adherence/mean': 0.910714328289032, 'rewards/count_adherence/std': 0.10628911107778549, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.39125001430511475, 'rewards/judge_quality/std': 0.2290625274181366, 'rewards/repeat_penalty/mean': 0.6733967065811157, 'rewards/repeat_penalty/std': 0.43340247869491577, 'rewards/near_duplicate_penalty/mean': 0.8702901601791382, 'rewards/near_duplicate_penalty/std': 0.29662787914276123, 'rewards/opening_diversity/mean': 0.887499988079071, 'rewards/opening_diversity/std': 0.3181980550289154, 'rewards/distinct_2/mean': 0.6950549483299255, 'rewards/distinct_2/std': 0.43958988785743713, 'rewards/total_composite/mean': 0.2903161644935608, 'rewards/total_composite/std': 0.17332184314727783, 'reward': 0.2903161644935608, 'reward_std': 0.17332184314727783, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1199302077293396, 'sampling/sampling_logp_difference/max': 2.8319387435913086, 'sampling/importance_sampling_ratio/min': 0.05889855697751045, 'sampling/importance_sampling_ratio/mean': 1.0077662467956543, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6010359786450863, 'clip_ratio/low_mean': 0.04199825134128332, 'clip_ratio/low_min': 0.04199825134128332, 'clip_ratio/high_mean': 0.0649514002725482, 'clip_ratio/high_max': 0.0649514002725482, 'clip_ratio/region_mean': 0.10694965161383152, 'reward_total_mean': 0.2903161644935608, 'reward_meter_mean': 0.801067590713501, 'reward_meter_std': 0.18090085685253143, 'reward_count_adherence_mean': 0.910714328289032, 'reward_count_adherence_std': 0.10628911107778549, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6733967065811157, 'reward_repeat_penalty_std': 0.43340247869491577, 'reward_near_duplicate_penalty_mean': 0.8702901601791382, 'reward_near_duplicate_penalty_std': 0.29662787914276123, 'reward_opening_diversity_mean': 0.887499988079071, 'reward_opening_diversity_std': 0.3181980550289154, 'reward_distinct_2_mean': 0.6950549483299255, 'reward_distinct_2_std': 0.43958988785743713, 'reward_judge_quality_mean': 0.39125001430511475, 'reward_judge_quality_std': 0.2290625274181366, 'reward_total_composite_mean': 0.2903161644935608, 'reward_total_composite_std': 0.17332184314727783, 'epoch': 0.02}

 16%|█▌        | 517/3300 [2:01:02<11:03:04, 14.30s/it]INFO 04-12 12:47:14 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 518/3300 [2:01:15<10:55:27, 14.14s/it]2026-04-12 12:47:28,023 | INFO | train_grpo_train | metrics_logged mode=train step=518

                                                       
{'loss': -0.05, 'grad_norm': 5.1548848152160645, 'learning_rate': 8.433333333333334e-06, 'num_tokens': 1035857.0, 'completions/mean_length': 113.75, 'completions/min_length': 48.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 56.857147216796875, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.6722588539123535, 'rewards/meter/std': 0.35397157073020935, '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/lexical_plausibility/mean': 0.9697229862213135, 'rewards/lexical_plausibility/std': 0.056428972631692886, 'rewards/judge_quality/mean': 0.4087499976158142, 'rewards/judge_quality/std': 0.258370578289032, 'rewards/repeat_penalty/mean': 0.9477458000183105, 'rewards/repeat_penalty/std': 0.08248929679393768, 'rewards/near_duplicate_penalty/mean': 0.9768604636192322, 'rewards/near_duplicate_penalty/std': 0.03228720650076866, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9687514305114746, 'rewards/distinct_2/std': 0.057349421083927155, 'rewards/total_composite/mean': 0.30740541219711304, 'rewards/total_composite/std': 0.26373177766799927, 'reward': 0.30740541219711304, 'reward_std': 0.26373177766799927, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17935426533222198, 'sampling/sampling_logp_difference/max': 1.8109140396118164, 'sampling/importance_sampling_ratio/min': 0.16350461542606354, 'sampling/importance_sampling_ratio/mean': 1.0112996101379395, 'sampling/importance_sampling_ratio/max': 1.8057483434677124, 'entropy': 1.397463046014309, 'clip_ratio/low_mean': 0.03603896126151085, 'clip_ratio/low_min': 0.03603896126151085, 'clip_ratio/high_mean': 0.09854637738317251, 'clip_ratio/high_max': 0.09854637738317251, 'clip_ratio/region_mean': 0.13458533864468336, 'reward_total_mean': 0.30740541219711304, 'reward_meter_mean': 0.6722588539123535, 'reward_meter_std': 0.35397157073020935, '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_lexical_plausibility_mean': 0.9697229862213135, 'reward_lexical_plausibility_std': 0.056428972631692886, 'reward_repeat_penalty_mean': 0.9477458000183105, 'reward_repeat_penalty_std': 0.08248929679393768, 'reward_near_duplicate_penalty_mean': 0.9768604636192322, 'reward_near_duplicate_penalty_std': 0.03228720650076866, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9687514305114746, 'reward_distinct_2_std': 0.057349421083927155, 'reward_judge_quality_mean': 0.4087499976158142, 'reward_judge_quality_std': 0.258370578289032, 'reward_total_composite_mean': 0.30740541219711304, 'reward_total_composite_std': 0.26373177766799927, 'epoch': 0.02}

 16%|█▌        | 518/3300 [2:01:15<10:55:27, 14.14s/it]INFO 04-12 12:47:28 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 519/3300 [2:01:26<10:07:40, 13.11s/it]2026-04-12 12:47:38,744 | INFO | train_grpo_train | metrics_logged mode=train step=519

                                                       
{'loss': 0.123, 'grad_norm': 11.445374488830566, 'learning_rate': 8.43030303030303e-06, 'num_tokens': 1037961.0, 'completions/mean_length': 93.0, 'completions/min_length': 77.0, 'completions/max_length': 124.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 93.0, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 124.0, 'rewards/meter/mean': 0.6855309009552002, 'rewards/meter/std': 0.40771737694740295, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.17251639068126678, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4375, 'rewards/judge_quality/std': 0.1642080545425415, 'rewards/repeat_penalty/mean': 0.9280847907066345, 'rewards/repeat_penalty/std': 0.07003556936979294, 'rewards/near_duplicate_penalty/mean': 0.9765260815620422, 'rewards/near_duplicate_penalty/std': 0.016959194093942642, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.9565455913543701, 'rewards/distinct_2/std': 0.0430552177131176, 'rewards/total_composite/mean': 0.22460846602916718, 'rewards/total_composite/std': 0.19476187229156494, 'reward': 0.22460846602916718, 'reward_std': 0.19476187229156494, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14586910605430603, 'sampling/sampling_logp_difference/max': 1.9536882638931274, 'sampling/importance_sampling_ratio/min': 0.1417502909898758, 'sampling/importance_sampling_ratio/mean': 1.0058424472808838, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8782564327120781, 'clip_ratio/low_mean': 0.07089105155318975, 'clip_ratio/low_min': 0.07089105155318975, 'clip_ratio/high_mean': 0.07607285212725401, 'clip_ratio/high_max': 0.07607285212725401, 'clip_ratio/region_mean': 0.14696390368044376, 'reward_total_mean': 0.22460846602916718, 'reward_meter_mean': 0.6855309009552002, 'reward_meter_std': 0.40771737694740295, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.17251639068126678, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9280847907066345, 'reward_repeat_penalty_std': 0.07003556936979294, 'reward_near_duplicate_penalty_mean': 0.9765260815620422, 'reward_near_duplicate_penalty_std': 0.016959194093942642, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.9565455913543701, 'reward_distinct_2_std': 0.0430552177131176, 'reward_judge_quality_mean': 0.4375, 'reward_judge_quality_std': 0.1642080545425415, 'reward_total_composite_mean': 0.22460846602916718, 'reward_total_composite_std': 0.19476187229156494, 'epoch': 0.02}

 16%|█▌        | 519/3300 [2:01:26<10:07:40, 13.11s/it]INFO 04-12 12:47:38 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 520/3300 [2:01:35<9:14:44, 11.97s/it] 2026-04-12 12:47:48,066 | INFO | train_grpo_train | metrics_logged mode=train step=520

                                                      
{'loss': -0.0109, 'grad_norm': 11.413832664489746, 'learning_rate': 8.427272727272729e-06, 'num_tokens': 1039813.0, 'completions/mean_length': 61.5, 'completions/min_length': 42.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.5, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.8347108364105225, 'rewards/meter/std': 0.2469102293252945, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6800000071525574, 'rewards/judge_quality/std': 0.24622870981693268, 'rewards/repeat_penalty/mean': 0.815872073173523, 'rewards/repeat_penalty/std': 0.26995933055877686, 'rewards/near_duplicate_penalty/mean': 0.9238532781600952, 'rewards/near_duplicate_penalty/std': 0.08535701036453247, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.8854270577430725, 'rewards/distinct_2/std': 0.19207239151000977, 'rewards/total_composite/mean': 0.5657850503921509, 'rewards/total_composite/std': 0.29925888776779175, 'reward': 0.5657850503921509, 'reward_std': 0.29925885796546936, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1263042390346527, 'sampling/sampling_logp_difference/max': 1.798166275024414, 'sampling/importance_sampling_ratio/min': 0.16560228168964386, 'sampling/importance_sampling_ratio/mean': 1.0113784074783325, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8484212383627892, 'clip_ratio/low_mean': 0.055868412367999554, 'clip_ratio/low_min': 0.055868412367999554, 'clip_ratio/high_mean': 0.04534399416297674, 'clip_ratio/high_max': 0.04534399416297674, 'clip_ratio/region_mean': 0.1012124065309763, 'reward_total_mean': 0.5657850503921509, 'reward_meter_mean': 0.8347108364105225, 'reward_meter_std': 0.2469102293252945, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.815872073173523, 'reward_repeat_penalty_std': 0.26995933055877686, 'reward_near_duplicate_penalty_mean': 0.9238532781600952, 'reward_near_duplicate_penalty_std': 0.08535701036453247, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.8854270577430725, 'reward_distinct_2_std': 0.19207239151000977, 'reward_judge_quality_mean': 0.6800000071525574, 'reward_judge_quality_std': 0.24622870981693268, 'reward_total_composite_mean': 0.5657850503921509, 'reward_total_composite_std': 0.29925888776779175, 'epoch': 0.02}

 16%|█▌        | 520/3300 [2:01:36<9:14:44, 11.97s/it]INFO 04-12 12:47:48 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 521/3300 [2:01:50<9:47:04, 12.68s/it]2026-04-12 12:48:02,374 | INFO | train_grpo_train | metrics_logged mode=train step=521

                                                      
{'loss': -0.1316, 'grad_norm': 4.403454780578613, 'learning_rate': 8.424242424242425e-06, 'num_tokens': 1043081.0, 'completions/mean_length': 244.5, 'completions/min_length': 180.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 206.2857208251953, 'completions/min_terminated_length': 180.0, 'completions/max_terminated_length': 239.0, 'rewards/meter/mean': 0.7262649536132812, 'rewards/meter/std': 0.23929134011268616, 'rewards/count_adherence/mean': 0.910714328289032, 'rewards/count_adherence/std': 0.07393559068441391, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9717724919319153, 'rewards/lexical_plausibility/std': 0.07983947545289993, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.16035674512386322, 'rewards/repeat_penalty/mean': 0.6779124736785889, 'rewards/repeat_penalty/std': 0.34433984756469727, 'rewards/near_duplicate_penalty/mean': 0.9364027976989746, 'rewards/near_duplicate_penalty/std': 0.062103621661663055, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.7290850281715393, 'rewards/distinct_2/std': 0.3030446469783783, 'rewards/total_composite/mean': 0.2239198386669159, 'rewards/total_composite/std': 0.16303850710391998, 'reward': 0.2239198386669159, 'reward_std': 0.1630384922027588, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13012371957302094, 'sampling/sampling_logp_difference/max': 2.064903974533081, 'sampling/importance_sampling_ratio/min': 0.12683045864105225, 'sampling/importance_sampling_ratio/mean': 1.003973126411438, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.848668709397316, 'clip_ratio/low_mean': 0.02434131968766451, 'clip_ratio/low_min': 0.02434131968766451, 'clip_ratio/high_mean': 0.09618020989000797, 'clip_ratio/high_max': 0.09618020989000797, 'clip_ratio/region_mean': 0.12052152957767248, 'reward_total_mean': 0.2239198386669159, 'reward_meter_mean': 0.7262649536132812, 'reward_meter_std': 0.23929134011268616, 'reward_count_adherence_mean': 0.910714328289032, 'reward_count_adherence_std': 0.07393559068441391, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9717724919319153, 'reward_lexical_plausibility_std': 0.07983947545289993, 'reward_repeat_penalty_mean': 0.6779124736785889, 'reward_repeat_penalty_std': 0.34433984756469727, 'reward_near_duplicate_penalty_mean': 0.9364027976989746, 'reward_near_duplicate_penalty_std': 0.062103621661663055, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.7290850281715393, 'reward_distinct_2_std': 0.3030446469783783, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.16035674512386322, 'reward_total_composite_mean': 0.2239198386669159, 'reward_total_composite_std': 0.16303850710391998, 'epoch': 0.02}

 16%|█▌        | 521/3300 [2:01:50<9:47:04, 12.68s/it]INFO 04-12 12:48:02 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 522/3300 [2:01:58<8:47:40, 11.40s/it]2026-04-12 12:48:10,792 | INFO | train_grpo_train | metrics_logged mode=train step=522

                                                      
{'loss': 0.071, 'grad_norm': 14.616438865661621, 'learning_rate': 8.421212121212122e-06, 'num_tokens': 1044721.0, 'completions/mean_length': 45.0, 'completions/min_length': 38.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.0, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.830040693283081, 'rewards/meter/std': 0.29204705357551575, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9866071343421936, 'rewards/lexical_plausibility/std': 0.03788072615861893, 'rewards/judge_quality/mean': 0.6725000143051147, 'rewards/judge_quality/std': 0.2666056752204895, 'rewards/repeat_penalty/mean': 0.9830596446990967, 'rewards/repeat_penalty/std': 0.02361476793885231, 'rewards/near_duplicate_penalty/mean': 0.9830596446990967, 'rewards/near_duplicate_penalty/std': 0.02361476793885231, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5864711999893188, 'rewards/total_composite/std': 0.3036585748195648, 'reward': 0.5864711999893188, 'reward_std': 0.30365854501724243, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14478832483291626, 'sampling/sampling_logp_difference/max': 1.0192508697509766, 'sampling/importance_sampling_ratio/min': 0.36086517572402954, 'sampling/importance_sampling_ratio/mean': 1.0219646692276, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1900311186909676, 'clip_ratio/low_mean': 0.0848024021834135, 'clip_ratio/low_min': 0.0848024021834135, 'clip_ratio/high_mean': 0.07249094545841217, 'clip_ratio/high_max': 0.07249094545841217, 'clip_ratio/region_mean': 0.15729334764182568, 'reward_total_mean': 0.5864711999893188, 'reward_meter_mean': 0.830040693283081, 'reward_meter_std': 0.29204705357551575, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9866071343421936, 'reward_lexical_plausibility_std': 0.03788072615861893, 'reward_repeat_penalty_mean': 0.9830596446990967, 'reward_repeat_penalty_std': 0.02361476793885231, 'reward_near_duplicate_penalty_mean': 0.9830596446990967, 'reward_near_duplicate_penalty_std': 0.02361476793885231, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6725000143051147, 'reward_judge_quality_std': 0.2666056752204895, 'reward_total_composite_mean': 0.5864711999893188, 'reward_total_composite_std': 0.3036585748195648, 'epoch': 0.02}

 16%|█▌        | 522/3300 [2:01:58<8:47:40, 11.40s/it]INFO 04-12 12:48:11 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 523/3300 [2:02:13<9:32:29, 12.37s/it]2026-04-12 12:48:25,428 | INFO | train_grpo_train | metrics_logged mode=train step=523

                                                      
{'loss': -0.0128, 'grad_norm': 3.289843797683716, 'learning_rate': 8.418181818181819e-06, 'num_tokens': 1046024.0, 'completions/mean_length': 143.875, 'completions/min_length': 15.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 21.166667938232422, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.775905191898346, 'rewards/meter/std': 0.3631601333618164, '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/lexical_plausibility/mean': 0.8998688459396362, 'rewards/lexical_plausibility/std': 0.1405652016401291, 'rewards/judge_quality/mean': 0.32124999165534973, 'rewards/judge_quality/std': 0.2923519015312195, 'rewards/repeat_penalty/mean': 0.7355769276618958, 'rewards/repeat_penalty/std': 0.04079463332891464, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9807692170143127, 'rewards/distinct_2/std': 0.054392825812101364, 'rewards/total_composite/mean': 0.2925552725791931, 'rewards/total_composite/std': 0.31374436616897583, 'reward': 0.2925552725791931, 'reward_std': 0.31374436616897583, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18874205648899078, 'sampling/sampling_logp_difference/max': 1.1223993301391602, 'sampling/importance_sampling_ratio/min': 0.3254978656768799, 'sampling/importance_sampling_ratio/mean': 1.0439214706420898, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5853483974933624, 'clip_ratio/low_mean': 0.030833333730697632, 'clip_ratio/low_min': 0.030833333730697632, 'clip_ratio/high_mean': 0.07590579893440008, 'clip_ratio/high_max': 0.07590579893440008, 'clip_ratio/region_mean': 0.10673913266509771, 'reward_total_mean': 0.2925552725791931, 'reward_meter_mean': 0.775905191898346, 'reward_meter_std': 0.3631601333618164, '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_lexical_plausibility_mean': 0.8998688459396362, 'reward_lexical_plausibility_std': 0.1405652016401291, 'reward_repeat_penalty_mean': 0.7355769276618958, 'reward_repeat_penalty_std': 0.04079463332891464, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9807692170143127, 'reward_distinct_2_std': 0.054392825812101364, 'reward_judge_quality_mean': 0.32124999165534973, 'reward_judge_quality_std': 0.2923519015312195, 'reward_total_composite_mean': 0.2925552725791931, 'reward_total_composite_std': 0.31374436616897583, 'epoch': 0.02}

 16%|█▌        | 523/3300 [2:02:13<9:32:29, 12.37s/it]INFO 04-12 12:48:25 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 524/3300 [2:02:28<10:17:08, 13.34s/it]2026-04-12 12:48:41,028 | INFO | train_grpo_train | metrics_logged mode=train step=524

                                                       
{'loss': -0.0733, 'grad_norm': 4.9495849609375, 'learning_rate': 8.415151515151516e-06, 'num_tokens': 1048289.0, 'completions/mean_length': 165.125, 'completions/min_length': 104.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 115.5714340209961, 'completions/min_terminated_length': 104.0, 'completions/max_terminated_length': 137.0, 'rewards/meter/mean': 0.7533771991729736, 'rewards/meter/std': 0.3445163667201996, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9759731888771057, 'rewards/lexical_plausibility/std': 0.06795810908079147, 'rewards/judge_quality/mean': 0.39625000953674316, 'rewards/judge_quality/std': 0.2585087716579437, 'rewards/repeat_penalty/mean': 0.9777988791465759, 'rewards/repeat_penalty/std': 0.022853916510939598, 'rewards/near_duplicate_penalty/mean': 0.9864957928657532, 'rewards/near_duplicate_penalty/std': 0.011978745460510254, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9911289215087891, 'rewards/distinct_2/std': 0.016462093219161034, 'rewards/total_composite/mean': 0.2796931862831116, 'rewards/total_composite/std': 0.2679268419742584, 'reward': 0.2796931862831116, 'reward_std': 0.2679268419742584, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16886699199676514, 'sampling/sampling_logp_difference/max': 3.5659232139587402, 'sampling/importance_sampling_ratio/min': 0.09188912063837051, 'sampling/importance_sampling_ratio/mean': 1.0322113037109375, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9957059249281883, 'clip_ratio/low_mean': 0.05708314571529627, 'clip_ratio/low_min': 0.05708314571529627, 'clip_ratio/high_mean': 0.0743851400911808, 'clip_ratio/high_max': 0.0743851400911808, 'clip_ratio/region_mean': 0.13146828580647707, 'reward_total_mean': 0.2796931862831116, 'reward_meter_mean': 0.7533771991729736, 'reward_meter_std': 0.3445163667201996, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9759731888771057, 'reward_lexical_plausibility_std': 0.06795810908079147, 'reward_repeat_penalty_mean': 0.9777988791465759, 'reward_repeat_penalty_std': 0.022853916510939598, 'reward_near_duplicate_penalty_mean': 0.9864957928657532, 'reward_near_duplicate_penalty_std': 0.011978745460510254, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9911289215087891, 'reward_distinct_2_std': 0.016462093219161034, 'reward_judge_quality_mean': 0.39625000953674316, 'reward_judge_quality_std': 0.2585087716579437, 'reward_total_composite_mean': 0.2796931862831116, 'reward_total_composite_std': 0.2679268419742584, 'epoch': 0.02}

 16%|█▌        | 524/3300 [2:02:29<10:17:08, 13.34s/it]INFO 04-12 12:48:41 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 525/3300 [2:02:40<9:48:32, 12.73s/it] 2026-04-12 12:48:52,323 | INFO | train_grpo_train | metrics_logged mode=train step=525

                                                      
{'loss': 0.0726, 'grad_norm': 9.855093002319336, 'learning_rate': 8.412121212121212e-06, 'num_tokens': 1050317.0, 'completions/mean_length': 92.5, 'completions/min_length': 64.0, 'completions/max_length': 127.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 92.5, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 127.0, 'rewards/meter/mean': 0.35597190260887146, 'rewards/meter/std': 0.30366161465644836, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.42499998211860657, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.9686890244483948, 'rewards/repeat_penalty/std': 0.02773984707891941, 'rewards/near_duplicate_penalty/mean': 0.9841871857643127, 'rewards/near_duplicate_penalty/std': 0.011527543887495995, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9841750860214233, 'rewards/distinct_2/std': 0.022115401923656464, 'rewards/total_composite/mean': 0.14699724316596985, 'rewards/total_composite/std': 0.11610212922096252, 'reward': 0.14699724316596985, 'reward_std': 0.11610212922096252, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17837126553058624, 'sampling/sampling_logp_difference/max': 1.7868213653564453, 'sampling/importance_sampling_ratio/min': 0.16749171912670135, 'sampling/importance_sampling_ratio/mean': 1.020738124847412, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3677703887224197, 'clip_ratio/low_mean': 0.06251568999141455, 'clip_ratio/low_min': 0.06251568999141455, 'clip_ratio/high_mean': 0.10365104116499424, 'clip_ratio/high_max': 0.10365104116499424, 'clip_ratio/region_mean': 0.1661667311564088, 'reward_total_mean': 0.14699724316596985, 'reward_meter_mean': 0.35597190260887146, 'reward_meter_std': 0.30366161465644836, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9686890244483948, 'reward_repeat_penalty_std': 0.02773984707891941, 'reward_near_duplicate_penalty_mean': 0.9841871857643127, 'reward_near_duplicate_penalty_std': 0.011527543887495995, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9841750860214233, 'reward_distinct_2_std': 0.022115401923656464, 'reward_judge_quality_mean': 0.42499998211860657, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.14699724316596985, 'reward_total_composite_std': 0.11610212922096252, 'epoch': 0.02}

 16%|█▌        | 525/3300 [2:02:40<9:48:32, 12.73s/it]INFO 04-12 12:48:52 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 526/3300 [2:02:49<9:04:15, 11.77s/it]2026-04-12 12:49:01,869 | INFO | train_grpo_train | metrics_logged mode=train step=526

                                                      
{'loss': 0.0512, 'grad_norm': 13.179574012756348, 'learning_rate': 8.40909090909091e-06, 'num_tokens': 1052070.0, 'completions/mean_length': 52.125, 'completions/min_length': 48.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.125, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.7835346460342407, 'rewards/meter/std': 0.2722063958644867, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9743589162826538, 'rewards/lexical_plausibility/std': 0.05855080857872963, 'rewards/judge_quality/mean': 0.7900000214576721, 'rewards/judge_quality/std': 0.24219238758087158, 'rewards/repeat_penalty/mean': 0.9872859716415405, 'rewards/repeat_penalty/std': 0.019436556845903397, 'rewards/near_duplicate_penalty/mean': 0.9872859716415405, 'rewards/near_duplicate_penalty/std': 0.019436556845903397, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.597822904586792, 'rewards/total_composite/std': 0.2706298828125, 'reward': 0.597822904586792, 'reward_std': 0.2706298530101776, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14874498546123505, 'sampling/sampling_logp_difference/max': 1.7132418155670166, 'sampling/importance_sampling_ratio/min': 0.1802804172039032, 'sampling/importance_sampling_ratio/mean': 1.0163997411727905, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.785020500421524, 'clip_ratio/low_mean': 0.08664832357317209, 'clip_ratio/low_min': 0.08664832357317209, 'clip_ratio/high_mean': 0.038968417793512344, 'clip_ratio/high_max': 0.038968417793512344, 'clip_ratio/region_mean': 0.12561674136668444, 'reward_total_mean': 0.597822904586792, 'reward_meter_mean': 0.7835346460342407, 'reward_meter_std': 0.2722063958644867, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9743589162826538, 'reward_lexical_plausibility_std': 0.05855080857872963, 'reward_repeat_penalty_mean': 0.9872859716415405, 'reward_repeat_penalty_std': 0.019436556845903397, 'reward_near_duplicate_penalty_mean': 0.9872859716415405, 'reward_near_duplicate_penalty_std': 0.019436556845903397, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7900000214576721, 'reward_judge_quality_std': 0.24219238758087158, 'reward_total_composite_mean': 0.597822904586792, 'reward_total_composite_std': 0.2706298828125, 'epoch': 0.02}

 16%|█▌        | 526/3300 [2:02:49<9:04:15, 11.77s/it]INFO 04-12 12:49:02 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 527/3300 [2:03:04<9:51:15, 12.79s/it]2026-04-12 12:49:17,048 | INFO | train_grpo_train | metrics_logged mode=train step=527

                                                      
{'loss': -0.0784, 'grad_norm': 2.7322332859039307, 'learning_rate': 8.406060606060606e-06, 'num_tokens': 1053736.0, 'completions/mean_length': 162.25, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 45.66666793823242, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.6106289029121399, 'rewards/meter/std': 0.3904173970222473, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9240602850914001, 'rewards/lexical_plausibility/std': 0.1270424872636795, 'rewards/judge_quality/mean': 0.33375000953674316, 'rewards/judge_quality/std': 0.28101539611816406, 'rewards/repeat_penalty/mean': 0.9891421794891357, 'rewards/repeat_penalty/std': 0.02679002285003662, 'rewards/near_duplicate_penalty/mean': 0.9993149638175964, 'rewards/near_duplicate_penalty/std': 0.0019375960109755397, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9898271560668945, 'rewards/distinct_2/std': 0.027016285806894302, 'rewards/total_composite/mean': 0.21866630017757416, 'rewards/total_composite/std': 0.17264395952224731, 'reward': 0.21866630017757416, 'reward_std': 0.17264395952224731, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17469477653503418, 'sampling/sampling_logp_difference/max': 1.054519772529602, 'sampling/importance_sampling_ratio/min': 0.37353944778442383, 'sampling/importance_sampling_ratio/mean': 1.03282630443573, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9298657774925232, 'clip_ratio/low_mean': 0.04841269925236702, 'clip_ratio/low_min': 0.04841269925236702, 'clip_ratio/high_mean': 0.09396970272064209, 'clip_ratio/high_max': 0.09396970272064209, 'clip_ratio/region_mean': 0.1423824019730091, 'reward_total_mean': 0.21866630017757416, 'reward_meter_mean': 0.6106289029121399, 'reward_meter_std': 0.3904173970222473, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9240602850914001, 'reward_lexical_plausibility_std': 0.1270424872636795, 'reward_repeat_penalty_mean': 0.9891421794891357, 'reward_repeat_penalty_std': 0.02679002285003662, 'reward_near_duplicate_penalty_mean': 0.9993149638175964, 'reward_near_duplicate_penalty_std': 0.0019375960109755397, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9898271560668945, 'reward_distinct_2_std': 0.027016285806894302, 'reward_judge_quality_mean': 0.33375000953674316, 'reward_judge_quality_std': 0.28101539611816406, 'reward_total_composite_mean': 0.21866630017757416, 'reward_total_composite_std': 0.17264395952224731, 'epoch': 0.02}

 16%|█▌        | 527/3300 [2:03:05<9:51:15, 12.79s/it]INFO 04-12 12:49:17 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 528/3300 [2:03:21<10:39:24, 13.84s/it]2026-04-12 12:49:33,331 | INFO | train_grpo_train | metrics_logged mode=train step=528

                                                       
{'loss': -0.1165, 'grad_norm': 4.148451805114746, 'learning_rate': 8.403030303030304e-06, 'num_tokens': 1056920.0, 'completions/mean_length': 250.0, 'completions/min_length': 186.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 212.57144165039062, 'completions/min_terminated_length': 186.0, 'completions/max_terminated_length': 307.0, 'rewards/meter/mean': 0.8554412126541138, 'rewards/meter/std': 0.18661826848983765, 'rewards/count_adherence/mean': 0.890625, 'rewards/count_adherence/std': 0.2158610224723816, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9729399681091309, 'rewards/lexical_plausibility/std': 0.07653743773698807, 'rewards/judge_quality/mean': 0.3162499964237213, 'rewards/judge_quality/std': 0.2650572657585144, 'rewards/repeat_penalty/mean': 0.4841121435165405, 'rewards/repeat_penalty/std': 0.451438844203949, 'rewards/near_duplicate_penalty/mean': 0.8259027004241943, 'rewards/near_duplicate_penalty/std': 0.1511022299528122, 'rewards/opening_diversity/mean': 0.8463541865348816, 'rewards/opening_diversity/std': 0.2396641969680786, 'rewards/distinct_2/mean': 0.5584592819213867, 'rewards/distinct_2/std': 0.4091425836086273, 'rewards/total_composite/mean': 0.2517433166503906, 'rewards/total_composite/std': 0.18851564824581146, 'reward': 0.2517433166503906, 'reward_std': 0.18851564824581146, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1041996031999588, 'sampling/sampling_logp_difference/max': 2.1392531394958496, 'sampling/importance_sampling_ratio/min': 0.11774274706840515, 'sampling/importance_sampling_ratio/mean': 1.0030522346496582, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5083982348442078, 'clip_ratio/low_mean': 0.03600514028221369, 'clip_ratio/low_min': 0.03600514028221369, 'clip_ratio/high_mean': 0.06670314725488424, 'clip_ratio/high_max': 0.06670314725488424, 'clip_ratio/region_mean': 0.10270828753709793, 'reward_total_mean': 0.2517433166503906, 'reward_meter_mean': 0.8554412126541138, 'reward_meter_std': 0.18661826848983765, 'reward_count_adherence_mean': 0.890625, 'reward_count_adherence_std': 0.2158610224723816, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9729399681091309, 'reward_lexical_plausibility_std': 0.07653743773698807, 'reward_repeat_penalty_mean': 0.4841121435165405, 'reward_repeat_penalty_std': 0.451438844203949, 'reward_near_duplicate_penalty_mean': 0.8259027004241943, 'reward_near_duplicate_penalty_std': 0.1511022299528122, 'reward_opening_diversity_mean': 0.8463541865348816, 'reward_opening_diversity_std': 0.2396641969680786, 'reward_distinct_2_mean': 0.5584592819213867, 'reward_distinct_2_std': 0.4091425836086273, 'reward_judge_quality_mean': 0.3162499964237213, 'reward_judge_quality_std': 0.2650572657585144, 'reward_total_composite_mean': 0.2517433166503906, 'reward_total_composite_std': 0.18851564824581146, 'epoch': 0.02}

 16%|█▌        | 528/3300 [2:03:21<10:39:24, 13.84s/it]INFO 04-12 12:49:33 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 529/3300 [2:03:36<10:59:33, 14.28s/it]2026-04-12 12:49:48,638 | INFO | train_grpo_train | metrics_logged mode=train step=529

                                                       
{'loss': -0.0277, 'grad_norm': 5.870815753936768, 'learning_rate': 8.400000000000001e-06, 'num_tokens': 1058633.0, 'completions/mean_length': 112.125, 'completions/min_length': 47.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 55.000003814697266, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.937852680683136, 'rewards/meter/std': 0.0940370187163353, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.8069376349449158, 'rewards/repeat_penalty/std': 0.3418090343475342, 'rewards/near_duplicate_penalty/mean': 0.9675962924957275, 'rewards/near_duplicate_penalty/std': 0.027440514415502548, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.8673076629638672, 'rewards/distinct_2/std': 0.3511062264442444, 'rewards/total_composite/mean': 0.299264520406723, 'rewards/total_composite/std': 0.15438225865364075, 'reward': 0.299264520406723, 'reward_std': 0.15438224375247955, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1524091511964798, 'sampling/sampling_logp_difference/max': 1.8996385335922241, 'sampling/importance_sampling_ratio/min': 0.14962269365787506, 'sampling/importance_sampling_ratio/mean': 1.0123872756958008, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9553992077708244, 'clip_ratio/low_mean': 0.03307629842311144, 'clip_ratio/low_min': 0.03307629842311144, 'clip_ratio/high_mean': 0.06976372841745615, 'clip_ratio/high_max': 0.06976372841745615, 'clip_ratio/region_mean': 0.10284002684056759, 'reward_total_mean': 0.299264520406723, 'reward_meter_mean': 0.937852680683136, 'reward_meter_std': 0.0940370187163353, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8069376349449158, 'reward_repeat_penalty_std': 0.3418090343475342, 'reward_near_duplicate_penalty_mean': 0.9675962924957275, 'reward_near_duplicate_penalty_std': 0.027440514415502548, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.8673076629638672, 'reward_distinct_2_std': 0.3511062264442444, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.299264520406723, 'reward_total_composite_std': 0.15438225865364075, 'epoch': 0.02}

 16%|█▌        | 529/3300 [2:03:36<10:59:33, 14.28s/it]INFO 04-12 12:49:48 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 530/3300 [2:03:48<10:28:38, 13.62s/it]2026-04-12 12:50:00,704 | INFO | train_grpo_train | metrics_logged mode=train step=530

                                                       
{'loss': 0.0651, 'grad_norm': 6.3402934074401855, 'learning_rate': 8.396969696969698e-06, 'num_tokens': 1062073.0, 'completions/mean_length': 190.0, 'completions/min_length': 167.0, 'completions/max_length': 207.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 190.0, 'completions/min_terminated_length': 167.0, 'completions/max_terminated_length': 207.0, 'rewards/meter/mean': 0.8997592926025391, 'rewards/meter/std': 0.1447831243276596, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0578637570142746, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9831730723381042, 'rewards/lexical_plausibility/std': 0.04759372025728226, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.08864051848649979, 'rewards/repeat_penalty/mean': 0.1761268973350525, 'rewards/repeat_penalty/std': 0.14410898089408875, 'rewards/near_duplicate_penalty/mean': 0.7250598669052124, 'rewards/near_duplicate_penalty/std': 0.08609069883823395, 'rewards/opening_diversity/mean': 0.8411458134651184, 'rewards/opening_diversity/std': 0.12284223735332489, 'rewards/distinct_2/mean': 0.3209337592124939, 'rewards/distinct_2/std': 0.2357674092054367, 'rewards/total_composite/mean': 0.2428385615348816, 'rewards/total_composite/std': 0.09923245012760162, 'reward': 0.2428385615348816, 'reward_std': 0.09923245012760162, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09381035715341568, 'sampling/sampling_logp_difference/max': 2.728348970413208, 'sampling/importance_sampling_ratio/min': 0.06532705575227737, 'sampling/importance_sampling_ratio/mean': 1.0126867294311523, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5182583890855312, 'clip_ratio/low_mean': 0.033622329123318195, 'clip_ratio/low_min': 0.033622329123318195, 'clip_ratio/high_mean': 0.05906510725617409, 'clip_ratio/high_max': 0.05906510725617409, 'clip_ratio/region_mean': 0.09268743637949228, 'reward_total_mean': 0.2428385615348816, 'reward_meter_mean': 0.8997592926025391, 'reward_meter_std': 0.1447831243276596, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0578637570142746, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9831730723381042, 'reward_lexical_plausibility_std': 0.04759372025728226, 'reward_repeat_penalty_mean': 0.1761268973350525, 'reward_repeat_penalty_std': 0.14410898089408875, 'reward_near_duplicate_penalty_mean': 0.7250598669052124, 'reward_near_duplicate_penalty_std': 0.08609069883823395, 'reward_opening_diversity_mean': 0.8411458134651184, 'reward_opening_diversity_std': 0.12284223735332489, 'reward_distinct_2_mean': 0.3209337592124939, 'reward_distinct_2_std': 0.2357674092054367, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.08864051848649979, 'reward_total_composite_mean': 0.2428385615348816, 'reward_total_composite_std': 0.09923245012760162, 'epoch': 0.02}

 16%|█▌        | 530/3300 [2:03:48<10:28:38, 13.62s/it]INFO 04-12 12:50:00 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 531/3300 [2:04:03<10:39:14, 13.85s/it]2026-04-12 12:50:15,105 | INFO | train_grpo_train | metrics_logged mode=train step=531

                                                       
{'loss': -0.1987, 'grad_norm': 2.5517168045043945, 'learning_rate': 8.393939393939394e-06, 'num_tokens': 1065203.0, 'completions/mean_length': 258.25, 'completions/min_length': 187.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 222.00001525878906, 'completions/min_terminated_length': 187.0, 'completions/max_terminated_length': 272.0, 'rewards/meter/mean': 0.8568861484527588, 'rewards/meter/std': 0.32454073429107666, 'rewards/count_adherence/mean': 0.796875, 'rewards/count_adherence/std': 0.27497971057891846, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9716575145721436, 'rewards/lexical_plausibility/std': 0.08016465604305267, 'rewards/judge_quality/mean': 0.13125000894069672, 'rewards/judge_quality/std': 0.0530330128967762, 'rewards/repeat_penalty/mean': 0.125, 'rewards/repeat_penalty/std': 0.3535533845424652, 'rewards/near_duplicate_penalty/mean': 0.5829541087150574, 'rewards/near_duplicate_penalty/std': 0.24616171419620514, 'rewards/opening_diversity/mean': 0.688725471496582, 'rewards/opening_diversity/std': 0.3646610975265503, 'rewards/distinct_2/mean': 0.125, 'rewards/distinct_2/std': 0.3535533845424652, 'rewards/total_composite/mean': 0.0985783189535141, 'rewards/total_composite/std': 0.05917447432875633, 'reward': 0.0985783189535141, 'reward_std': 0.059174466878175735, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04312387481331825, 'sampling/sampling_logp_difference/max': 1.377960205078125, 'sampling/importance_sampling_ratio/min': 0.25209224224090576, 'sampling/importance_sampling_ratio/mean': 1.0071038007736206, 'sampling/importance_sampling_ratio/max': 1.8820263147354126, 'entropy': 0.2851339690387249, 'clip_ratio/low_mean': 0.0065104165114462376, 'clip_ratio/low_min': 0.0065104165114462376, 'clip_ratio/high_mean': 0.029803105629980564, 'clip_ratio/high_max': 0.029803105629980564, 'clip_ratio/region_mean': 0.0363135221414268, 'reward_total_mean': 0.0985783189535141, 'reward_meter_mean': 0.8568861484527588, 'reward_meter_std': 0.32454073429107666, 'reward_count_adherence_mean': 0.796875, 'reward_count_adherence_std': 0.27497971057891846, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9716575145721436, 'reward_lexical_plausibility_std': 0.08016465604305267, 'reward_repeat_penalty_mean': 0.125, 'reward_repeat_penalty_std': 0.3535533845424652, 'reward_near_duplicate_penalty_mean': 0.5829541087150574, 'reward_near_duplicate_penalty_std': 0.24616171419620514, 'reward_opening_diversity_mean': 0.688725471496582, 'reward_opening_diversity_std': 0.3646610975265503, 'reward_distinct_2_mean': 0.125, 'reward_distinct_2_std': 0.3535533845424652, 'reward_judge_quality_mean': 0.13125000894069672, 'reward_judge_quality_std': 0.0530330128967762, 'reward_total_composite_mean': 0.0985783189535141, 'reward_total_composite_std': 0.05917447432875633, 'epoch': 0.02}

 16%|█▌        | 531/3300 [2:04:03<10:39:14, 13.85s/it]INFO 04-12 12:50:15 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 532/3300 [2:04:17<10:43:53, 13.96s/it]2026-04-12 12:50:29,313 | INFO | train_grpo_train | metrics_logged mode=train step=532

                                                       
{'loss': -0.0601, 'grad_norm': 5.555056095123291, 'learning_rate': 8.390909090909091e-06, 'num_tokens': 1067370.0, 'completions/mean_length': 162.875, 'completions/min_length': 93.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 113.00000762939453, 'completions/min_terminated_length': 93.0, 'completions/max_terminated_length': 135.0, 'rewards/meter/mean': 0.8084428310394287, 'rewards/meter/std': 0.24196060001850128, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4462500214576721, 'rewards/judge_quality/std': 0.2173829823732376, 'rewards/repeat_penalty/mean': 0.6832889318466187, 'rewards/repeat_penalty/std': 0.3724896311759949, 'rewards/near_duplicate_penalty/mean': 0.9544548988342285, 'rewards/near_duplicate_penalty/std': 0.0558350495994091, 'rewards/opening_diversity/mean': 0.9765625, 'rewards/opening_diversity/std': 0.032346826046705246, 'rewards/distinct_2/mean': 0.7420210838317871, 'rewards/distinct_2/std': 0.35255300998687744, 'rewards/total_composite/mean': 0.32075899839401245, 'rewards/total_composite/std': 0.20925991237163544, 'reward': 0.32075899839401245, 'reward_std': 0.20925991237163544, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14588786661624908, 'sampling/sampling_logp_difference/max': 1.3479371070861816, 'sampling/importance_sampling_ratio/min': 0.2597756087779999, 'sampling/importance_sampling_ratio/mean': 1.021937608718872, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9303793534636497, 'clip_ratio/low_mean': 0.03686714172363281, 'clip_ratio/low_min': 0.03686714172363281, 'clip_ratio/high_mean': 0.09394550323486328, 'clip_ratio/high_max': 0.09394550323486328, 'clip_ratio/region_mean': 0.1308126449584961, 'reward_total_mean': 0.32075899839401245, 'reward_meter_mean': 0.8084428310394287, 'reward_meter_std': 0.24196060001850128, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.12938730418682098, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6832889318466187, 'reward_repeat_penalty_std': 0.3724896311759949, 'reward_near_duplicate_penalty_mean': 0.9544548988342285, 'reward_near_duplicate_penalty_std': 0.0558350495994091, 'reward_opening_diversity_mean': 0.9765625, 'reward_opening_diversity_std': 0.032346826046705246, 'reward_distinct_2_mean': 0.7420210838317871, 'reward_distinct_2_std': 0.35255300998687744, 'reward_judge_quality_mean': 0.4462500214576721, 'reward_judge_quality_std': 0.2173829823732376, 'reward_total_composite_mean': 0.32075899839401245, 'reward_total_composite_std': 0.20925991237163544, 'epoch': 0.02}

 16%|█▌        | 532/3300 [2:04:17<10:43:53, 13.96s/it]INFO 04-12 12:50:29 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 533/3300 [2:04:32<10:56:19, 14.23s/it]2026-04-12 12:50:44,179 | INFO | train_grpo_train | metrics_logged mode=train step=533

                                                       
{'loss': -0.1584, 'grad_norm': 0.9167531132698059, 'learning_rate': 8.387878787878788e-06, 'num_tokens': 1070403.0, 'completions/mean_length': 489.125, 'completions/min_length': 436.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.625, 'completions/mean_terminated_length': 451.0, 'completions/min_terminated_length': 436.0, 'completions/max_terminated_length': 461.0, 'rewards/meter/mean': 0.8364267945289612, 'rewards/meter/std': 0.22743096947669983, 'rewards/count_adherence/mean': 0.7083333730697632, 'rewards/count_adherence/std': 0.13772527873516083, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.038550376892089844, 'rewards/repeat_penalty/std': 0.07170351594686508, 'rewards/near_duplicate_penalty/mean': 0.7515722513198853, 'rewards/near_duplicate_penalty/std': 0.1594141721725464, 'rewards/opening_diversity/mean': 0.9142134189605713, 'rewards/opening_diversity/std': 0.09840895235538483, 'rewards/distinct_2/mean': 0.05562567338347435, 'rewards/distinct_2/std': 0.10356443375349045, 'rewards/total_composite/mean': 0.09171570837497711, 'rewards/total_composite/std': 0.03053947724401951, 'reward': 0.09171570837497711, 'reward_std': 0.03053947724401951, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06979882717132568, 'sampling/sampling_logp_difference/max': 1.7257180213928223, 'sampling/importance_sampling_ratio/min': 0.17804516851902008, 'sampling/importance_sampling_ratio/mean': 1.0062392950057983, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.15940506011247635, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.023210786283016205, 'clip_ratio/high_max': 0.023210786283016205, 'clip_ratio/region_mean': 0.023210786283016205, 'reward_total_mean': 0.09171570837497711, 'reward_meter_mean': 0.8364267945289612, 'reward_meter_std': 0.22743096947669983, 'reward_count_adherence_mean': 0.7083333730697632, 'reward_count_adherence_std': 0.13772527873516083, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.038550376892089844, 'reward_repeat_penalty_std': 0.07170351594686508, 'reward_near_duplicate_penalty_mean': 0.7515722513198853, 'reward_near_duplicate_penalty_std': 0.1594141721725464, 'reward_opening_diversity_mean': 0.9142134189605713, 'reward_opening_diversity_std': 0.09840895235538483, 'reward_distinct_2_mean': 0.05562567338347435, 'reward_distinct_2_std': 0.10356443375349045, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.09171570837497711, 'reward_total_composite_std': 0.03053947724401951, 'epoch': 0.02}

 16%|█▌        | 533/3300 [2:04:32<10:56:19, 14.23s/it]INFO 04-12 12:50:44 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 534/3300 [2:04:42<10:05:40, 13.14s/it]2026-04-12 12:50:54,771 | INFO | train_grpo_train | metrics_logged mode=train step=534

                                                       
{'loss': 0.0342, 'grad_norm': 12.688547134399414, 'learning_rate': 8.384848484848485e-06, 'num_tokens': 1072694.0, 'completions/mean_length': 109.375, 'completions/min_length': 96.0, 'completions/max_length': 119.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 109.375, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 119.0, 'rewards/meter/mean': 0.7278497219085693, 'rewards/meter/std': 0.3690914809703827, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9895833730697632, 'rewards/lexical_plausibility/std': 0.029462775215506554, 'rewards/judge_quality/mean': 0.42499998211860657, 'rewards/judge_quality/std': 0.11649646610021591, 'rewards/repeat_penalty/mean': 0.7939637899398804, 'rewards/repeat_penalty/std': 0.29377561807632446, 'rewards/near_duplicate_penalty/mean': 0.9472005367279053, 'rewards/near_duplicate_penalty/std': 0.07469348609447479, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.8238227367401123, 'rewards/distinct_2/std': 0.27867135405540466, 'rewards/total_composite/mean': 0.285033643245697, 'rewards/total_composite/std': 0.13481025397777557, 'reward': 0.285033643245697, 'reward_std': 0.13481023907661438, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13370047509670258, 'sampling/sampling_logp_difference/max': 1.7642860412597656, 'sampling/importance_sampling_ratio/min': 0.17130905389785767, 'sampling/importance_sampling_ratio/mean': 1.0064051151275635, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9834525212645531, 'clip_ratio/low_mean': 0.07096594572067261, 'clip_ratio/low_min': 0.07096594572067261, 'clip_ratio/high_mean': 0.06541966227814555, 'clip_ratio/high_max': 0.06541966227814555, 'clip_ratio/region_mean': 0.13638560799881816, 'reward_total_mean': 0.285033643245697, 'reward_meter_mean': 0.7278497219085693, 'reward_meter_std': 0.3690914809703827, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9895833730697632, 'reward_lexical_plausibility_std': 0.029462775215506554, 'reward_repeat_penalty_mean': 0.7939637899398804, 'reward_repeat_penalty_std': 0.29377561807632446, 'reward_near_duplicate_penalty_mean': 0.9472005367279053, 'reward_near_duplicate_penalty_std': 0.07469348609447479, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.8238227367401123, 'reward_distinct_2_std': 0.27867135405540466, 'reward_judge_quality_mean': 0.42499998211860657, 'reward_judge_quality_std': 0.11649646610021591, 'reward_total_composite_mean': 0.285033643245697, 'reward_total_composite_std': 0.13481025397777557, 'epoch': 0.02}

 16%|█▌        | 534/3300 [2:04:42<10:05:40, 13.14s/it]INFO 04-12 12:50:55 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 535/3300 [2:04:53<9:26:23, 12.29s/it] 2026-04-12 12:51:05,083 | INFO | train_grpo_train | metrics_logged mode=train step=535

                                                      
{'loss': 0.0652, 'grad_norm': 14.839488983154297, 'learning_rate': 8.381818181818183e-06, 'num_tokens': 1075104.0, 'completions/mean_length': 103.25, 'completions/min_length': 85.0, 'completions/max_length': 119.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 103.25, 'completions/min_terminated_length': 85.0, 'completions/max_terminated_length': 119.0, 'rewards/meter/mean': 0.5845506191253662, 'rewards/meter/std': 0.15791787207126617, 'rewards/count_adherence/mean': 0.9583333134651184, 'rewards/count_adherence/std': 0.07715168595314026, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.36249998211860657, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.948292076587677, 'rewards/repeat_penalty/std': 0.046413879841566086, 'rewards/near_duplicate_penalty/mean': 0.9824104905128479, 'rewards/near_duplicate_penalty/std': 0.009192436933517456, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9650310277938843, 'rewards/distinct_2/std': 0.04034508392214775, 'rewards/total_composite/mean': 0.20369039475917816, 'rewards/total_composite/std': 0.09125176072120667, 'reward': 0.20369039475917816, 'reward_std': 0.09125176072120667, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21394693851470947, 'sampling/sampling_logp_difference/max': 2.5198817253112793, 'sampling/importance_sampling_ratio/min': 0.08046912401914597, 'sampling/importance_sampling_ratio/mean': 0.9883208870887756, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0349463522434235, 'clip_ratio/low_mean': 0.11953691951930523, 'clip_ratio/low_min': 0.11953691951930523, 'clip_ratio/high_mean': 0.08050814270973206, 'clip_ratio/high_max': 0.08050814270973206, 'clip_ratio/region_mean': 0.20004506222903728, 'reward_total_mean': 0.20369039475917816, 'reward_meter_mean': 0.5845506191253662, 'reward_meter_std': 0.15791787207126617, 'reward_count_adherence_mean': 0.9583333134651184, 'reward_count_adherence_std': 0.07715168595314026, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.948292076587677, 'reward_repeat_penalty_std': 0.046413879841566086, 'reward_near_duplicate_penalty_mean': 0.9824104905128479, 'reward_near_duplicate_penalty_std': 0.009192436933517456, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9650310277938843, 'reward_distinct_2_std': 0.04034508392214775, 'reward_judge_quality_mean': 0.36249998211860657, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.20369039475917816, 'reward_total_composite_std': 0.09125176072120667, 'epoch': 0.02}

 16%|█▌        | 535/3300 [2:04:53<9:26:23, 12.29s/it]INFO 04-12 12:51:05 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 536/3300 [2:05:09<10:20:14, 13.46s/it]2026-04-12 12:51:21,286 | INFO | train_grpo_train | metrics_logged mode=train step=536

                                                       
{'loss': -0.0424, 'grad_norm': 4.612396240234375, 'learning_rate': 8.37878787878788e-06, 'num_tokens': 1076626.0, 'completions/mean_length': 167.25, 'completions/min_length': 43.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 52.333335876464844, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.19621151685714722, 'rewards/meter/std': 0.24942757189273834, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9493439197540283, 'rewards/lexical_plausibility/std': 0.0961630567908287, 'rewards/judge_quality/mean': 0.5262500047683716, 'rewards/judge_quality/std': 0.3537932336330414, 'rewards/repeat_penalty/mean': 0.9899824857711792, 'rewards/repeat_penalty/std': 0.020527975633740425, 'rewards/near_duplicate_penalty/mean': 0.9899824857711792, 'rewards/near_duplicate_penalty/std': 0.020527975633740425, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1278495043516159, 'rewards/total_composite/std': 0.22685647010803223, 'reward': 0.1278495043516159, 'reward_std': 0.22685645520687103, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14579251408576965, 'sampling/sampling_logp_difference/max': 1.9719133377075195, 'sampling/importance_sampling_ratio/min': 0.13919028639793396, 'sampling/importance_sampling_ratio/mean': 1.0111684799194336, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7191209271550179, 'clip_ratio/low_mean': 0.08018747624009848, 'clip_ratio/low_min': 0.08018747624009848, 'clip_ratio/high_mean': 0.055055659264326096, 'clip_ratio/high_max': 0.055055659264326096, 'clip_ratio/region_mean': 0.13524313550442457, 'reward_total_mean': 0.1278495043516159, 'reward_meter_mean': 0.19621151685714722, 'reward_meter_std': 0.24942757189273834, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9493439197540283, 'reward_lexical_plausibility_std': 0.0961630567908287, 'reward_repeat_penalty_mean': 0.9899824857711792, 'reward_repeat_penalty_std': 0.020527975633740425, 'reward_near_duplicate_penalty_mean': 0.9899824857711792, 'reward_near_duplicate_penalty_std': 0.020527975633740425, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5262500047683716, 'reward_judge_quality_std': 0.3537932336330414, 'reward_total_composite_mean': 0.1278495043516159, 'reward_total_composite_std': 0.22685647010803223, 'epoch': 0.02}

 16%|█▌        | 536/3300 [2:05:09<10:20:14, 13.46s/it]INFO 04-12 12:51:21 [block_pool.py:378] Successfully reset prefix cache

 16%|█▋        | 537/3300 [2:05:17<9:14:54, 12.05s/it] 2026-04-12 12:51:30,035 | INFO | train_grpo_train | metrics_logged mode=train step=537

                                                      
{'loss': 0.0479, 'grad_norm': 12.68325138092041, 'learning_rate': 8.375757575757576e-06, 'num_tokens': 1078299.0, 'completions/mean_length': 49.125, 'completions/min_length': 44.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.125, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.8141142129898071, 'rewards/meter/std': 0.3222385346889496, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5674999952316284, 'rewards/judge_quality/std': 0.21756774187088013, 'rewards/repeat_penalty/mean': 0.9850510358810425, 'rewards/repeat_penalty/std': 0.029591161757707596, 'rewards/near_duplicate_penalty/mean': 0.9936339259147644, 'rewards/near_duplicate_penalty/std': 0.007461874280124903, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9912587404251099, 'rewards/distinct_2/std': 0.024724015966057777, 'rewards/total_composite/mean': 0.4584875702857971, 'rewards/total_composite/std': 0.23430661857128143, 'reward': 0.4584875702857971, 'reward_std': 0.23430663347244263, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1474178284406662, 'sampling/sampling_logp_difference/max': 1.5713319778442383, 'sampling/importance_sampling_ratio/min': 0.2077682614326477, 'sampling/importance_sampling_ratio/mean': 1.0108470916748047, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1057326719164848, 'clip_ratio/low_mean': 0.11327622085809708, 'clip_ratio/low_min': 0.11327622085809708, 'clip_ratio/high_mean': 0.040294526144862175, 'clip_ratio/high_max': 0.040294526144862175, 'clip_ratio/region_mean': 0.15357074700295925, 'reward_total_mean': 0.4584875702857971, 'reward_meter_mean': 0.8141142129898071, 'reward_meter_std': 0.3222385346889496, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9850510358810425, 'reward_repeat_penalty_std': 0.029591161757707596, 'reward_near_duplicate_penalty_mean': 0.9936339259147644, 'reward_near_duplicate_penalty_std': 0.007461874280124903, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9912587404251099, 'reward_distinct_2_std': 0.024724015966057777, 'reward_judge_quality_mean': 0.5674999952316284, 'reward_judge_quality_std': 0.21756774187088013, 'reward_total_composite_mean': 0.4584875702857971, 'reward_total_composite_std': 0.23430661857128143, 'epoch': 0.02}

 16%|█▋        | 537/3300 [2:05:18<9:14:54, 12.05s/it]INFO 04-12 12:51:30 [block_pool.py:378] Successfully reset prefix cache

 16%|█▋        | 538/3300 [2:05:29<9:04:56, 11.84s/it]2026-04-12 12:51:41,390 | INFO | train_grpo_train | metrics_logged mode=train step=538

                                                      
{'loss': 0.578, 'grad_norm': 11.540745735168457, 'learning_rate': 8.372727272727273e-06, 'num_tokens': 1079982.0, 'completions/mean_length': 54.375, 'completions/min_length': 20.0, 'completions/max_length': 275.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.375, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 275.0, 'rewards/meter/mean': 0.489854633808136, 'rewards/meter/std': 0.500199556350708, '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/lexical_plausibility/mean': 0.9736684560775757, 'rewards/lexical_plausibility/std': 0.053863029927015305, 'rewards/judge_quality/mean': 0.7150000333786011, 'rewards/judge_quality/std': 0.38052597641944885, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4499676525592804, 'rewards/total_composite/std': 0.4609576165676117, 'reward': 0.4499676525592804, 'reward_std': 0.4609576165676117, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14810021221637726, 'sampling/sampling_logp_difference/max': 1.21683931350708, 'sampling/importance_sampling_ratio/min': 0.29616478085517883, 'sampling/importance_sampling_ratio/mean': 1.0322062969207764, 'sampling/importance_sampling_ratio/max': 1.8346823453903198, 'entropy': 1.885888610035181, 'clip_ratio/low_mean': 0.05073616676963866, 'clip_ratio/low_min': 0.05073616676963866, 'clip_ratio/high_mean': 0.07121865265071392, 'clip_ratio/high_max': 0.07121865265071392, 'clip_ratio/region_mean': 0.12195481942035258, 'reward_total_mean': 0.4499676525592804, 'reward_meter_mean': 0.489854633808136, 'reward_meter_std': 0.500199556350708, '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_lexical_plausibility_mean': 0.9736684560775757, 'reward_lexical_plausibility_std': 0.053863029927015305, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7150000333786011, 'reward_judge_quality_std': 0.38052597641944885, 'reward_total_composite_mean': 0.4499676525592804, 'reward_total_composite_std': 0.4609576165676117, 'epoch': 0.02}

 16%|█▋        | 538/3300 [2:05:29<9:04:56, 11.84s/it]INFO 04-12 12:51:41 [block_pool.py:378] Successfully reset prefix cache

 16%|█▋        | 539/3300 [2:05:39<8:44:17, 11.39s/it]2026-04-12 12:51:51,732 | INFO | train_grpo_train | metrics_logged mode=train step=539

                                                      
{'loss': 0.0096, 'grad_norm': 8.234294891357422, 'learning_rate': 8.36969696969697e-06, 'num_tokens': 1082383.0, 'completions/mean_length': 111.125, 'completions/min_length': 101.0, 'completions/max_length': 128.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 111.125, 'completions/min_terminated_length': 101.0, 'completions/max_terminated_length': 128.0, 'rewards/meter/mean': 0.7686270475387573, 'rewards/meter/std': 0.3404479920864105, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.4210032820701599, 'rewards/repeat_penalty/std': 0.252521276473999, 'rewards/near_duplicate_penalty/mean': 0.8123122453689575, 'rewards/near_duplicate_penalty/std': 0.11966996639966965, 'rewards/opening_diversity/mean': 0.8250000476837158, 'rewards/opening_diversity/std': 0.20701968669891357, 'rewards/distinct_2/mean': 0.597417414188385, 'rewards/distinct_2/std': 0.187087744474411, 'rewards/total_composite/mean': 0.25443488359451294, 'rewards/total_composite/std': 0.14483973383903503, 'reward': 0.25443488359451294, 'reward_std': 0.14483973383903503, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1166553646326065, 'sampling/sampling_logp_difference/max': 1.492222785949707, 'sampling/importance_sampling_ratio/min': 0.22487226128578186, 'sampling/importance_sampling_ratio/mean': 1.024922251701355, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7824344448745251, 'clip_ratio/low_mean': 0.027238872833549976, 'clip_ratio/low_min': 0.027238872833549976, 'clip_ratio/high_mean': 0.07031228579580784, 'clip_ratio/high_max': 0.07031228579580784, 'clip_ratio/region_mean': 0.09755115862935781, 'reward_total_mean': 0.25443488359451294, 'reward_meter_mean': 0.7686270475387573, 'reward_meter_std': 0.3404479920864105, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.4210032820701599, 'reward_repeat_penalty_std': 0.252521276473999, 'reward_near_duplicate_penalty_mean': 0.8123122453689575, 'reward_near_duplicate_penalty_std': 0.11966996639966965, 'reward_opening_diversity_mean': 0.8250000476837158, 'reward_opening_diversity_std': 0.20701968669891357, 'reward_distinct_2_mean': 0.597417414188385, 'reward_distinct_2_std': 0.187087744474411, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.25443488359451294, 'reward_total_composite_std': 0.14483973383903503, 'epoch': 0.02}

 16%|█▋        | 539/3300 [2:05:39<8:44:17, 11.39s/it]INFO 04-12 12:51:51 [block_pool.py:378] Successfully reset prefix cache

 16%|█▋        | 540/3300 [2:05:53<9:22:50, 12.24s/it]2026-04-12 12:52:05,937 | INFO | train_grpo_train | metrics_logged mode=train step=540

                                                      
{'loss': -0.0678, 'grad_norm': 5.527726173400879, 'learning_rate': 8.366666666666667e-06, 'num_tokens': 1084091.0, 'completions/mean_length': 106.5, 'completions/min_length': 45.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 48.57143020629883, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.7789685130119324, 'rewards/meter/std': 0.3174363374710083, '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/lexical_plausibility/mean': 0.9723135232925415, 'rewards/lexical_plausibility/std': 0.07830909639596939, 'rewards/judge_quality/mean': 0.5162500143051147, 'rewards/judge_quality/std': 0.2506527006626129, 'rewards/repeat_penalty/mean': 0.9699653387069702, 'rewards/repeat_penalty/std': 0.050749704241752625, 'rewards/near_duplicate_penalty/mean': 0.98801189661026, 'rewards/near_duplicate_penalty/std': 0.01910880208015442, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9812270998954773, 'rewards/distinct_2/std': 0.034774426370859146, 'rewards/total_composite/mean': 0.4167323708534241, 'rewards/total_composite/std': 0.29724326729774475, 'reward': 0.4167323708534241, 'reward_std': 0.29724326729774475, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16650508344173431, 'sampling/sampling_logp_difference/max': 1.545924186706543, 'sampling/importance_sampling_ratio/min': 0.21311481297016144, 'sampling/importance_sampling_ratio/mean': 1.0367190837860107, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0137456208467484, 'clip_ratio/low_mean': 0.04594863811507821, 'clip_ratio/low_min': 0.04594863811507821, 'clip_ratio/high_mean': 0.07743352837860584, 'clip_ratio/high_max': 0.07743352837860584, 'clip_ratio/region_mean': 0.12338216649368405, 'reward_total_mean': 0.4167323708534241, 'reward_meter_mean': 0.7789685130119324, 'reward_meter_std': 0.3174363374710083, '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_lexical_plausibility_mean': 0.9723135232925415, 'reward_lexical_plausibility_std': 0.07830909639596939, 'reward_repeat_penalty_mean': 0.9699653387069702, 'reward_repeat_penalty_std': 0.050749704241752625, 'reward_near_duplicate_penalty_mean': 0.98801189661026, 'reward_near_duplicate_penalty_std': 0.01910880208015442, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9812270998954773, 'reward_distinct_2_std': 0.034774426370859146, 'reward_judge_quality_mean': 0.5162500143051147, 'reward_judge_quality_std': 0.2506527006626129, 'reward_total_composite_mean': 0.4167323708534241, 'reward_total_composite_std': 0.29724326729774475, 'epoch': 0.02}

 16%|█▋        | 540/3300 [2:05:53<9:22:50, 12.24s/it]INFO 04-12 12:52:06 [block_pool.py:378] Successfully reset prefix cache

 16%|█▋        | 541/3300 [2:06:02<8:38:48, 11.28s/it]2026-04-12 12:52:14,997 | INFO | train_grpo_train | metrics_logged mode=train step=541

                                                      
{'loss': 0.0907, 'grad_norm': 13.243159294128418, 'learning_rate': 8.363636363636365e-06, 'num_tokens': 1085833.0, 'completions/mean_length': 50.75, 'completions/min_length': 45.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.75, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.7803934812545776, 'rewards/meter/std': 0.3405003547668457, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.4337500035762787, 'rewards/judge_quality/std': 0.21999594569206238, 'rewards/repeat_penalty/mean': 0.9894441962242126, 'rewards/repeat_penalty/std': 0.019719798117876053, 'rewards/near_duplicate_penalty/mean': 0.9894441962242126, 'rewards/near_duplicate_penalty/std': 0.019719798117876053, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3628407418727875, 'rewards/total_composite/std': 0.26674890518188477, 'reward': 0.3628407418727875, 'reward_std': 0.2667488753795624, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16147242486476898, 'sampling/sampling_logp_difference/max': 1.7120890617370605, 'sampling/importance_sampling_ratio/min': 0.18048834800720215, 'sampling/importance_sampling_ratio/mean': 1.030494213104248, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9272258132696152, 'clip_ratio/low_mean': 0.06402011448517442, 'clip_ratio/low_min': 0.06402011448517442, 'clip_ratio/high_mean': 0.056826923973858356, 'clip_ratio/high_max': 0.056826923973858356, 'clip_ratio/region_mean': 0.12084703845903277, 'reward_total_mean': 0.3628407418727875, 'reward_meter_mean': 0.7803934812545776, 'reward_meter_std': 0.3405003547668457, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.9894441962242126, 'reward_repeat_penalty_std': 0.019719798117876053, 'reward_near_duplicate_penalty_mean': 0.9894441962242126, 'reward_near_duplicate_penalty_std': 0.019719798117876053, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4337500035762787, 'reward_judge_quality_std': 0.21999594569206238, 'reward_total_composite_mean': 0.3628407418727875, 'reward_total_composite_std': 0.26674890518188477, 'epoch': 0.02}

 16%|█▋        | 541/3300 [2:06:02<8:38:48, 11.28s/it]INFO 04-12 12:52:15 [block_pool.py:378] Successfully reset prefix cache

 16%|█▋        | 542/3300 [2:06:17<9:20:06, 12.19s/it]2026-04-12 12:52:29,291 | INFO | train_grpo_train | metrics_logged mode=train step=542

                                                      
{'loss': -0.0194, 'grad_norm': 5.271066665649414, 'learning_rate': 8.360606060606062e-06, 'num_tokens': 1087513.0, 'completions/mean_length': 110.0, 'completions/min_length': 47.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 52.57143020629883, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.5333980917930603, 'rewards/meter/std': 0.4215831160545349, '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/lexical_plausibility/mean': 0.9752260446548462, 'rewards/lexical_plausibility/std': 0.07007137686014175, 'rewards/judge_quality/mean': 0.6000000238418579, 'rewards/judge_quality/std': 0.34050384163856506, 'rewards/repeat_penalty/mean': 0.8951529860496521, 'rewards/repeat_penalty/std': 0.2170160710811615, 'rewards/near_duplicate_penalty/mean': 0.9672882556915283, 'rewards/near_duplicate_penalty/std': 0.040903035551309586, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.953177273273468, 'rewards/distinct_2/std': 0.10797474533319473, 'rewards/total_composite/mean': 0.3980540931224823, 'rewards/total_composite/std': 0.3673470914363861, 'reward': 0.3980540931224823, 'reward_std': 0.3673470914363861, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15299922227859497, 'sampling/sampling_logp_difference/max': 1.871394395828247, 'sampling/importance_sampling_ratio/min': 0.15390890836715698, 'sampling/importance_sampling_ratio/mean': 1.0171422958374023, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0272081568837166, 'clip_ratio/low_mean': 0.035638296976685524, 'clip_ratio/low_min': 0.035638296976685524, 'clip_ratio/high_mean': 0.0830766474828124, 'clip_ratio/high_max': 0.0830766474828124, 'clip_ratio/region_mean': 0.11871494445949793, 'reward_total_mean': 0.3980540931224823, 'reward_meter_mean': 0.5333980917930603, 'reward_meter_std': 0.4215831160545349, '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_lexical_plausibility_mean': 0.9752260446548462, 'reward_lexical_plausibility_std': 0.07007137686014175, 'reward_repeat_penalty_mean': 0.8951529860496521, 'reward_repeat_penalty_std': 0.2170160710811615, 'reward_near_duplicate_penalty_mean': 0.9672882556915283, 'reward_near_duplicate_penalty_std': 0.040903035551309586, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.953177273273468, 'reward_distinct_2_std': 0.10797474533319473, 'reward_judge_quality_mean': 0.6000000238418579, 'reward_judge_quality_std': 0.34050384163856506, 'reward_total_composite_mean': 0.3980540931224823, 'reward_total_composite_std': 0.3673470914363861, 'epoch': 0.02}

 16%|█▋        | 542/3300 [2:06:17<9:20:06, 12.19s/it]INFO 04-12 12:52:29 [block_pool.py:378] Successfully reset prefix cache

 16%|█▋        | 543/3300 [2:06:31<9:53:07, 12.91s/it]2026-04-12 12:52:43,883 | INFO | train_grpo_train | metrics_logged mode=train step=543

                                                      
{'loss': -0.0798, 'grad_norm': 5.470287322998047, 'learning_rate': 8.357575757575759e-06, 'num_tokens': 1089265.0, 'completions/mean_length': 126.0, 'completions/min_length': 65.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 70.85714721679688, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.6617354154586792, 'rewards/meter/std': 0.25543513894081116, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9628397226333618, 'rewards/lexical_plausibility/std': 0.1051052063703537, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.1414213478565216, 'rewards/repeat_penalty/mean': 0.9738050699234009, 'rewards/repeat_penalty/std': 0.04037882015109062, 'rewards/near_duplicate_penalty/mean': 0.9850949645042419, 'rewards/near_duplicate_penalty/std': 0.013675507158041, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9883450269699097, 'rewards/distinct_2/std': 0.03296534717082977, 'rewards/total_composite/mean': 0.24361778795719147, 'rewards/total_composite/std': 0.1561136245727539, 'reward': 0.24361778795719147, 'reward_std': 0.1561136096715927, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16222620010375977, 'sampling/sampling_logp_difference/max': 1.6206461191177368, 'sampling/importance_sampling_ratio/min': 0.1977708786725998, 'sampling/importance_sampling_ratio/mean': 0.9903451204299927, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3137017339468002, 'clip_ratio/low_mean': 0.06880779284983873, 'clip_ratio/low_min': 0.06880779284983873, 'clip_ratio/high_mean': 0.0771570224314928, 'clip_ratio/high_max': 0.0771570224314928, 'clip_ratio/region_mean': 0.14596481528133154, 'reward_total_mean': 0.24361778795719147, 'reward_meter_mean': 0.6617354154586792, 'reward_meter_std': 0.25543513894081116, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9628397226333618, 'reward_lexical_plausibility_std': 0.1051052063703537, 'reward_repeat_penalty_mean': 0.9738050699234009, 'reward_repeat_penalty_std': 0.04037882015109062, 'reward_near_duplicate_penalty_mean': 0.9850949645042419, 'reward_near_duplicate_penalty_std': 0.013675507158041, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9883450269699097, 'reward_distinct_2_std': 0.03296534717082977, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.1414213478565216, 'reward_total_composite_mean': 0.24361778795719147, 'reward_total_composite_std': 0.1561136245727539, 'epoch': 0.02}

 16%|█▋        | 543/3300 [2:06:31<9:53:07, 12.91s/it]INFO 04-12 12:52:44 [block_pool.py:378] Successfully reset prefix cache

 16%|█▋        | 544/3300 [2:06:49<11:05:35, 14.49s/it]2026-04-12 12:53:02,064 | INFO | train_grpo_train | metrics_logged mode=train step=544

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 8.354545454545455e-06, 'num_tokens': 1091265.0, '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': 0.9187072515487671, 'rewards/meter/std': 0.09112033993005753, 'rewards/count_adherence/mean': 0.9342105388641357, 'rewards/count_adherence/std': 0.06745999306440353, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.16250000894069672, 'rewards/judge_quality/std': 0.0353553369641304, 'rewards/repeat_penalty/mean': 0.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 0.4728279113769531, 'rewards/near_duplicate_penalty/std': 0.13525299727916718, 'rewards/opening_diversity/mean': 0.6840425133705139, 'rewards/opening_diversity/std': 0.21817994117736816, 'rewards/distinct_2/mean': 0.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.14112301170825958, 'rewards/total_composite/std': 0.04260226711630821, 'reward': 0.14112301170825958, 'reward_std': 0.042602263391017914, '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': 0.14112301170825958, 'reward_meter_mean': 0.9187072515487671, 'reward_meter_std': 0.09112033993005753, 'reward_count_adherence_mean': 0.9342105388641357, 'reward_count_adherence_std': 0.06745999306440353, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.0, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 0.4728279113769531, 'reward_near_duplicate_penalty_std': 0.13525299727916718, 'reward_opening_diversity_mean': 0.6840425133705139, 'reward_opening_diversity_std': 0.21817994117736816, 'reward_distinct_2_mean': 0.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.16250000894069672, 'reward_judge_quality_std': 0.0353553369641304, 'reward_total_composite_mean': 0.14112301170825958, 'reward_total_composite_std': 0.04260226711630821, 'epoch': 0.02}

 16%|█▋        | 544/3300 [2:06:50<11:05:35, 14.49s/it]INFO 04-12 12:53:02 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 545/3300 [2:07:00<10:07:16, 13.23s/it]2026-04-12 12:53:12,338 | INFO | train_grpo_train | metrics_logged mode=train step=545

                                                       
{'loss': 0.0616, 'grad_norm': 13.080174446105957, 'learning_rate': 8.351515151515152e-06, 'num_tokens': 1092848.0, 'completions/mean_length': 41.875, 'completions/min_length': 37.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.875, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.6356033086776733, 'rewards/meter/std': 0.3513253331184387, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9866071343421936, 'rewards/lexical_plausibility/std': 0.03788072615861893, 'rewards/judge_quality/mean': 0.706250011920929, 'rewards/judge_quality/std': 0.3091896176338196, 'rewards/repeat_penalty/mean': 0.9848418235778809, 'rewards/repeat_penalty/std': 0.013436457142233849, 'rewards/near_duplicate_penalty/mean': 0.9848418235778809, 'rewards/near_duplicate_penalty/std': 0.013436457142233849, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4831591546535492, 'rewards/total_composite/std': 0.3251233994960785, 'reward': 0.4831591546535492, 'reward_std': 0.3251233994960785, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.162849560379982, 'sampling/sampling_logp_difference/max': 1.457371711730957, 'sampling/importance_sampling_ratio/min': 0.2328474521636963, 'sampling/importance_sampling_ratio/mean': 1.0163511037826538, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1603094786405563, 'clip_ratio/low_mean': 0.09493415802717209, 'clip_ratio/low_min': 0.09493415802717209, 'clip_ratio/high_mean': 0.050484295934438705, 'clip_ratio/high_max': 0.050484295934438705, 'clip_ratio/region_mean': 0.1454184539616108, 'reward_total_mean': 0.4831591546535492, 'reward_meter_mean': 0.6356033086776733, 'reward_meter_std': 0.3513253331184387, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9866071343421936, 'reward_lexical_plausibility_std': 0.03788072615861893, 'reward_repeat_penalty_mean': 0.9848418235778809, 'reward_repeat_penalty_std': 0.013436457142233849, 'reward_near_duplicate_penalty_mean': 0.9848418235778809, 'reward_near_duplicate_penalty_std': 0.013436457142233849, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.706250011920929, 'reward_judge_quality_std': 0.3091896176338196, 'reward_total_composite_mean': 0.4831591546535492, 'reward_total_composite_std': 0.3251233994960785, 'epoch': 0.02}

 17%|█▋        | 545/3300 [2:07:00<10:07:16, 13.23s/it]INFO 04-12 12:53:12 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 546/3300 [2:07:15<10:28:29, 13.69s/it]2026-04-12 12:53:27,118 | INFO | train_grpo_train | metrics_logged mode=train step=546

                                                       
{'loss': -0.0296, 'grad_norm': 6.41250467300415, 'learning_rate': 8.348484848484849e-06, 'num_tokens': 1094288.0, 'completions/mean_length': 103.0, 'completions/min_length': 41.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 44.57143020629883, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.4175654649734497, 'rewards/meter/std': 0.4242997169494629, '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/lexical_plausibility/mean': 0.9850408434867859, 'rewards/lexical_plausibility/std': 0.042310841381549835, 'rewards/judge_quality/mean': 0.5600000619888306, 'rewards/judge_quality/std': 0.39402684569358826, 'rewards/repeat_penalty/mean': 0.8592712879180908, 'rewards/repeat_penalty/std': 0.25859907269477844, 'rewards/near_duplicate_penalty/mean': 0.960735559463501, 'rewards/near_duplicate_penalty/std': 0.06616660207509995, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.952729344367981, 'rewards/distinct_2/std': 0.1289743185043335, 'rewards/total_composite/mean': 0.2499169409275055, 'rewards/total_composite/std': 0.36938366293907166, 'reward': 0.2499169409275055, 'reward_std': 0.36938366293907166, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1582486778497696, 'sampling/sampling_logp_difference/max': 2.2328763008117676, 'sampling/importance_sampling_ratio/min': 0.10721959173679352, 'sampling/importance_sampling_ratio/mean': 1.0251611471176147, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9267575144767761, 'clip_ratio/low_mean': 0.07908797264099121, 'clip_ratio/low_min': 0.07908797264099121, 'clip_ratio/high_mean': 0.03888889029622078, 'clip_ratio/high_max': 0.03888889029622078, 'clip_ratio/region_mean': 0.11797686293721199, 'reward_total_mean': 0.2499169409275055, 'reward_meter_mean': 0.4175654649734497, 'reward_meter_std': 0.4242997169494629, '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_lexical_plausibility_mean': 0.9850408434867859, 'reward_lexical_plausibility_std': 0.042310841381549835, 'reward_repeat_penalty_mean': 0.8592712879180908, 'reward_repeat_penalty_std': 0.25859907269477844, 'reward_near_duplicate_penalty_mean': 0.960735559463501, 'reward_near_duplicate_penalty_std': 0.06616660207509995, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.952729344367981, 'reward_distinct_2_std': 0.1289743185043335, 'reward_judge_quality_mean': 0.5600000619888306, 'reward_judge_quality_std': 0.39402684569358826, 'reward_total_composite_mean': 0.2499169409275055, 'reward_total_composite_std': 0.36938366293907166, 'epoch': 0.02}

 17%|█▋        | 546/3300 [2:07:15<10:28:29, 13.69s/it]INFO 04-12 12:53:27 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 547/3300 [2:07:29<10:38:55, 13.93s/it]2026-04-12 12:53:41,585 | INFO | train_grpo_train | metrics_logged mode=train step=547

                                                       
{'loss': -0.1237, 'grad_norm': 4.194162368774414, 'learning_rate': 8.345454545454546e-06, 'num_tokens': 1096002.0, 'completions/mean_length': 114.25, 'completions/min_length': 52.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 57.42857360839844, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.777793824672699, 'rewards/meter/std': 0.320039838552475, '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/lexical_plausibility/mean': 0.9824602603912354, 'rewards/lexical_plausibility/std': 0.0496099591255188, 'rewards/judge_quality/mean': 0.5049999952316284, 'rewards/judge_quality/std': 0.28972893953323364, 'rewards/repeat_penalty/mean': 0.8999760150909424, 'rewards/repeat_penalty/std': 0.17061437666416168, 'rewards/near_duplicate_penalty/mean': 0.9790429472923279, 'rewards/near_duplicate_penalty/std': 0.03674924373626709, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.971382737159729, 'rewards/distinct_2/std': 0.05943568795919418, 'rewards/total_composite/mean': 0.3156241774559021, 'rewards/total_composite/std': 0.1547367125749588, 'reward': 0.3156241774559021, 'reward_std': 0.1547366827726364, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14931762218475342, 'sampling/sampling_logp_difference/max': 1.6100870370864868, 'sampling/importance_sampling_ratio/min': 0.19987021386623383, 'sampling/importance_sampling_ratio/mean': 1.0040839910507202, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9953689277172089, 'clip_ratio/low_mean': 0.0190677959471941, 'clip_ratio/low_min': 0.0190677959471941, 'clip_ratio/high_mean': 0.1263645440340042, 'clip_ratio/high_max': 0.1263645440340042, 'clip_ratio/region_mean': 0.1454323399811983, 'reward_total_mean': 0.3156241774559021, 'reward_meter_mean': 0.777793824672699, 'reward_meter_std': 0.320039838552475, '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_lexical_plausibility_mean': 0.9824602603912354, 'reward_lexical_plausibility_std': 0.0496099591255188, 'reward_repeat_penalty_mean': 0.8999760150909424, 'reward_repeat_penalty_std': 0.17061437666416168, 'reward_near_duplicate_penalty_mean': 0.9790429472923279, 'reward_near_duplicate_penalty_std': 0.03674924373626709, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.971382737159729, 'reward_distinct_2_std': 0.05943568795919418, 'reward_judge_quality_mean': 0.5049999952316284, 'reward_judge_quality_std': 0.28972893953323364, 'reward_total_composite_mean': 0.3156241774559021, 'reward_total_composite_std': 0.1547367125749588, 'epoch': 0.02}

 17%|█▋        | 547/3300 [2:07:29<10:38:55, 13.93s/it]INFO 04-12 12:53:41 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 548/3300 [2:07:38<9:25:09, 12.32s/it] 2026-04-12 12:53:50,172 | INFO | train_grpo_train | metrics_logged mode=train step=548

                                                      
{'loss': 0.0532, 'grad_norm': 15.030717849731445, 'learning_rate': 8.342424242424244e-06, 'num_tokens': 1097629.0, 'completions/mean_length': 46.375, 'completions/min_length': 39.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.375, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.8358132243156433, 'rewards/meter/std': 0.2915712296962738, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6087499856948853, 'rewards/judge_quality/std': 0.24930977821350098, 'rewards/repeat_penalty/mean': 0.9405306577682495, 'rewards/repeat_penalty/std': 0.15187403559684753, 'rewards/near_duplicate_penalty/mean': 0.9794235825538635, 'rewards/near_duplicate_penalty/std': 0.04282815009355545, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9825174808502197, 'rewards/distinct_2/std': 0.049448031932115555, 'rewards/total_composite/mean': 0.48486417531967163, 'rewards/total_composite/std': 0.2695642411708832, 'reward': 0.48486417531967163, 'reward_std': 0.2695642411708832, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16091519594192505, 'sampling/sampling_logp_difference/max': 2.071469783782959, 'sampling/importance_sampling_ratio/min': 0.12600044906139374, 'sampling/importance_sampling_ratio/mean': 1.0105230808258057, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2028623595833778, 'clip_ratio/low_mean': 0.09606942720711231, 'clip_ratio/low_min': 0.09606942720711231, 'clip_ratio/high_mean': 0.04661835730075836, 'clip_ratio/high_max': 0.04661835730075836, 'clip_ratio/region_mean': 0.14268778450787067, 'reward_total_mean': 0.48486417531967163, 'reward_meter_mean': 0.8358132243156433, 'reward_meter_std': 0.2915712296962738, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9405306577682495, 'reward_repeat_penalty_std': 0.15187403559684753, 'reward_near_duplicate_penalty_mean': 0.9794235825538635, 'reward_near_duplicate_penalty_std': 0.04282815009355545, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9825174808502197, 'reward_distinct_2_std': 0.049448031932115555, 'reward_judge_quality_mean': 0.6087499856948853, 'reward_judge_quality_std': 0.24930977821350098, 'reward_total_composite_mean': 0.48486417531967163, 'reward_total_composite_std': 0.2695642411708832, 'epoch': 0.02}

 17%|█▋        | 548/3300 [2:07:38<9:25:09, 12.32s/it]INFO 04-12 12:53:50 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 549/3300 [2:07:52<9:54:30, 12.97s/it]2026-04-12 12:54:04,636 | INFO | train_grpo_train | metrics_logged mode=train step=549

                                                      
{'loss': -0.0953, 'grad_norm': 1.6906237602233887, 'learning_rate': 8.339393939393941e-06, 'num_tokens': 1099190.0, 'completions/mean_length': 349.125, 'completions/min_length': 72.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.625, 'completions/mean_terminated_length': 77.66667175292969, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.4822614789009094, 'rewards/meter/std': 0.342801034450531, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.17251639068126678, 'rewards/arabic_clean/mean': 0.375, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.8447681665420532, 'rewards/lexical_plausibility/std': 0.13131478428840637, 'rewards/judge_quality/mean': 0.25874999165534973, 'rewards/judge_quality/std': 0.30898162722587585, 'rewards/repeat_penalty/mean': 0.9817433953285217, 'rewards/repeat_penalty/std': 0.0381956622004509, 'rewards/near_duplicate_penalty/mean': 0.993170976638794, 'rewards/near_duplicate_penalty/std': 0.009107104502618313, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9883450269699097, 'rewards/distinct_2/std': 0.03296534717082977, 'rewards/total_composite/mean': 0.09634813666343689, 'rewards/total_composite/std': 0.1397472321987152, 'reward': 0.09634813666343689, 'reward_std': 0.13974721729755402, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1432182788848877, 'sampling/sampling_logp_difference/max': 1.026468276977539, 'sampling/importance_sampling_ratio/min': 0.3582700490951538, 'sampling/importance_sampling_ratio/mean': 1.0189342498779297, 'sampling/importance_sampling_ratio/max': 1.7254616022109985, 'entropy': 0.4322194755077362, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.06151787377893925, 'clip_ratio/high_max': 0.06151787377893925, 'clip_ratio/region_mean': 0.06151787377893925, 'reward_total_mean': 0.09634813666343689, 'reward_meter_mean': 0.4822614789009094, 'reward_meter_std': 0.342801034450531, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.17251639068126678, 'reward_arabic_clean_mean': 0.375, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.8447681665420532, 'reward_lexical_plausibility_std': 0.13131478428840637, 'reward_repeat_penalty_mean': 0.9817433953285217, 'reward_repeat_penalty_std': 0.0381956622004509, 'reward_near_duplicate_penalty_mean': 0.993170976638794, 'reward_near_duplicate_penalty_std': 0.009107104502618313, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9883450269699097, 'reward_distinct_2_std': 0.03296534717082977, 'reward_judge_quality_mean': 0.25874999165534973, 'reward_judge_quality_std': 0.30898162722587585, 'reward_total_composite_mean': 0.09634813666343689, 'reward_total_composite_std': 0.1397472321987152, 'epoch': 0.02}

 17%|█▋        | 549/3300 [2:07:52<9:54:30, 12.97s/it]INFO 04-12 12:54:04 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 550/3300 [2:08:07<10:17:49, 13.48s/it]2026-04-12 12:54:19,317 | INFO | train_grpo_train | metrics_logged mode=train step=550

                                                       
{'loss': -0.0697, 'grad_norm': 6.1772990226745605, 'learning_rate': 8.336363636363636e-06, 'num_tokens': 1100997.0, 'completions/mean_length': 117.875, 'completions/min_length': 39.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 61.57143020629883, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.5366573333740234, 'rewards/meter/std': 0.2995971739292145, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.3563483655452728, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9671957492828369, 'rewards/lexical_plausibility/std': 0.0927843302488327, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.9931343197822571, 'rewards/repeat_penalty/std': 0.007398698944598436, 'rewards/near_duplicate_penalty/mean': 0.9931343197822571, 'rewards/near_duplicate_penalty/std': 0.007398698944598436, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.20977303385734558, 'rewards/total_composite/std': 0.11658559739589691, 'reward': 0.20977303385734558, 'reward_std': 0.11658558994531631, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17950139939785004, 'sampling/sampling_logp_difference/max': 1.9874672889709473, 'sampling/importance_sampling_ratio/min': 0.13972559571266174, 'sampling/importance_sampling_ratio/mean': 0.9971045255661011, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8207018300890923, 'clip_ratio/low_mean': 0.06258048303425312, 'clip_ratio/low_min': 0.06258048303425312, 'clip_ratio/high_mean': 0.07775097154080868, 'clip_ratio/high_max': 0.07775097154080868, 'clip_ratio/region_mean': 0.1403314545750618, 'reward_total_mean': 0.20977303385734558, 'reward_meter_mean': 0.5366573333740234, 'reward_meter_std': 0.2995971739292145, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.3563483655452728, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9671957492828369, 'reward_lexical_plausibility_std': 0.0927843302488327, 'reward_repeat_penalty_mean': 0.9931343197822571, 'reward_repeat_penalty_std': 0.007398698944598436, 'reward_near_duplicate_penalty_mean': 0.9931343197822571, 'reward_near_duplicate_penalty_std': 0.007398698944598436, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.20977303385734558, 'reward_total_composite_std': 0.11658559739589691, 'epoch': 0.02}

 17%|█▋        | 550/3300 [2:08:07<10:17:49, 13.48s/it]INFO 04-12 12:54:19 [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 [00:11<01:01,  5.64s/it]

 23%|██▎       | 3/13 [00:22<01:21,  8.14s/it]

 31%|███       | 4/13 [00:34<01:26,  9.58s/it]

 38%|███▊      | 5/13 [00:45<01:19,  9.96s/it]

 46%|████▌     | 6/13 [00:58<01:16, 10.86s/it]

 54%|█████▍    | 7/13 [01:10<01:07, 11.18s/it]

 62%|██████▏   | 8/13 [01:22<00:57, 11.59s/it]

 69%|██████▉   | 9/13 [01:34<00:46, 11.70s/it]

 77%|███████▋  | 10/13 [01:46<00:35, 11.81s/it]

 85%|████████▍ | 11/13 [01:58<00:23, 11.76s/it]

 92%|█████████▏| 12/13 [02:09<00:11, 11.49s/it]

100%|██████████| 13/13 [02:20<00:00, 11.47s/it]2026-04-12 12:56:51,524 | INFO | train_grpo_train | metrics_logged mode=eval step=550

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 152.1646, 'eval_samples_per_second': 0.683, 'eval_steps_per_second': 0.085, 'eval_num_tokens': 1100997.0, 'eval_completions/mean_length': 220.76923076923077, 'eval_completions/min_length': 43.84615384615385, 'eval_completions/max_length': 500.0769230769231, 'eval_completions/clipped_ratio': 0.16346153846153846, 'eval_completions/mean_terminated_length': 161.85723935640775, 'eval_completions/min_terminated_length': 43.84615384615385, 'eval_completions/max_terminated_length': 363.53846153846155, 'eval_rewards/meter/mean': 0.586402918283756, 'eval_rewards/meter/std': 0.36957008563555205, 'eval_rewards/count_adherence/mean': 0.8603702325087327, 'eval_rewards/count_adherence/std': 0.2073669834778859, 'eval_rewards/arabic_clean/mean': 0.8365384615384616, 'eval_rewards/arabic_clean/std': 0.2693372047864474, 'eval_rewards/lexical_plausibility/mean': 0.9697205057510963, 'eval_rewards/lexical_plausibility/std': 0.0657596132503106, 'eval_rewards/judge_quality/mean': 0.356730770606261, 'eval_rewards/judge_quality/std': 0.22836716014605302, 'eval_rewards/repeat_penalty/mean': 0.6280048993917612, 'eval_rewards/repeat_penalty/std': 0.410148391356835, 'eval_rewards/near_duplicate_penalty/mean': 0.8681904627726629, 'eval_rewards/near_duplicate_penalty/std': 0.18287275760219648, 'eval_rewards/opening_diversity/mean': 0.9148014187812805, 'eval_rewards/opening_diversity/std': 0.17345992848277092, 'eval_rewards/distinct_2/mean': 0.6751560202011695, 'eval_rewards/distinct_2/std': 0.41054011422854203, 'eval_rewards/total_composite/mean': 0.17228719764030898, 'eval_rewards/total_composite/std': 0.17016129195690155, 'eval_reward': 0.17228719764030898, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.05730216749585592, 'eval_sampling/sampling_logp_difference/max': 1.1263794532189002, 'eval_sampling/importance_sampling_ratio/min': 0.3301447331905365, 'eval_sampling/importance_sampling_ratio/mean': 1.0154129266738892, 'eval_sampling/importance_sampling_ratio/max': 1.5247919009282038, 'eval_entropy': 0.6868736079105964, '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': 0.17228719764030898, 'eval_reward_meter_mean': 0.586402918283756, 'eval_reward_meter_std': 0.36957008563555205, 'eval_reward_count_adherence_mean': 0.8603702325087327, 'eval_reward_count_adherence_std': 0.2073669834778859, 'eval_reward_arabic_clean_mean': 0.8365384615384616, 'eval_reward_arabic_clean_std': 0.2693372047864474, 'eval_reward_lexical_plausibility_mean': 0.9697205057510963, 'eval_reward_lexical_plausibility_std': 0.0657596132503106, 'eval_reward_repeat_penalty_mean': 0.6280048993917612, 'eval_reward_repeat_penalty_std': 0.410148391356835, 'eval_reward_near_duplicate_penalty_mean': 0.8681904627726629, 'eval_reward_near_duplicate_penalty_std': 0.18287275760219648, 'eval_reward_opening_diversity_mean': 0.9148014187812805, 'eval_reward_opening_diversity_std': 0.17345992848277092, 'eval_reward_distinct_2_mean': 0.6751560202011695, 'eval_reward_distinct_2_std': 0.41054011422854203, 'eval_reward_judge_quality_mean': 0.356730770606261, 'eval_reward_judge_quality_std': 0.22836716014605302, 'eval_reward_total_composite_mean': 0.17228719764030898, 'eval_reward_total_composite_std': 0.17016129195690155, 'epoch': 0.02}

 17%|█▋        | 550/3300 [2:10:39<10:17:49, 13.48s/it]

100%|██████████| 13/13 [02:20<00:00, 11.47s/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(

 17%|█▋        | 551/3300 [2:10:56<46:00:26, 60.25s/it]2026-04-12 12:57:08,696 | INFO | train_grpo_train | metrics_logged mode=train step=551

                                                       
{'loss': -0.153, 'grad_norm': 2.5289623737335205, 'learning_rate': 8.333333333333334e-06, 'num_tokens': 1103062.0, 'completions/mean_length': 322.125, 'completions/min_length': 116.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 132.25, 'completions/min_terminated_length': 116.0, 'completions/max_terminated_length': 150.0, 'rewards/meter/mean': 0.46613040566444397, 'rewards/meter/std': 0.4457698166370392, 'rewards/count_adherence/mean': 0.7250000238418579, 'rewards/count_adherence/std': 0.3011881411075592, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/lexical_plausibility/mean': 0.8561485409736633, 'rewards/lexical_plausibility/std': 0.15561705827713013, 'rewards/judge_quality/mean': 0.3087500035762787, 'rewards/judge_quality/std': 0.3165184557437897, 'rewards/repeat_penalty/mean': 0.9961950778961182, 'rewards/repeat_penalty/std': 0.004838768858462572, 'rewards/near_duplicate_penalty/mean': 0.9961950778961182, 'rewards/near_duplicate_penalty/std': 0.004838768858462572, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.24493418633937836, 'rewards/total_composite/std': 0.3308947682380676, 'reward': 0.24493418633937836, 'reward_std': 0.3308947682380676, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.181468665599823, 'sampling/sampling_logp_difference/max': 1.959695816040039, 'sampling/importance_sampling_ratio/min': 0.14090128242969513, 'sampling/importance_sampling_ratio/mean': 1.038630723953247, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.803390696644783, 'clip_ratio/low_mean': 0.02822580561041832, 'clip_ratio/low_min': 0.02822580561041832, 'clip_ratio/high_mean': 0.06060313992202282, 'clip_ratio/high_max': 0.06060313992202282, 'clip_ratio/region_mean': 0.08882894553244114, 'reward_total_mean': 0.24493418633937836, 'reward_meter_mean': 0.46613040566444397, 'reward_meter_std': 0.4457698166370392, 'reward_count_adherence_mean': 0.7250000238418579, 'reward_count_adherence_std': 0.3011881411075592, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_lexical_plausibility_mean': 0.8561485409736633, 'reward_lexical_plausibility_std': 0.15561705827713013, 'reward_repeat_penalty_mean': 0.9961950778961182, 'reward_repeat_penalty_std': 0.004838768858462572, 'reward_near_duplicate_penalty_mean': 0.9961950778961182, 'reward_near_duplicate_penalty_std': 0.004838768858462572, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3087500035762787, 'reward_judge_quality_std': 0.3165184557437897, 'reward_total_composite_mean': 0.24493418633937836, 'reward_total_composite_std': 0.3308947682380676, 'epoch': 0.02}

 17%|█▋        | 551/3300 [2:10:56<46:00:26, 60.25s/it]INFO 04-12 12:57:08 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 552/3300 [2:11:10<35:18:07, 46.25s/it]2026-04-12 12:57:22,275 | INFO | train_grpo_train | metrics_logged mode=train step=552

                                                       
{'loss': -0.0378, 'grad_norm': 2.1509552001953125, 'learning_rate': 8.330303030303031e-06, 'num_tokens': 1104403.0, 'completions/mean_length': 143.625, 'completions/min_length': 17.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 20.83333396911621, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.7479416728019714, 'rewards/meter/std': 0.4325871467590332, '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/lexical_plausibility/mean': 0.904184877872467, 'rewards/lexical_plausibility/std': 0.13383860886096954, 'rewards/judge_quality/mean': 0.35500001907348633, 'rewards/judge_quality/std': 0.36245197057724, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3364430069923401, 'rewards/total_composite/std': 0.3685567080974579, 'reward': 0.3364430069923401, 'reward_std': 0.3685566782951355, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15116764605045319, 'sampling/sampling_logp_difference/max': 1.1842036247253418, 'sampling/importance_sampling_ratio/min': 0.3059897720813751, 'sampling/importance_sampling_ratio/mean': 1.0206120014190674, 'sampling/importance_sampling_ratio/max': 1.8944429159164429, 'entropy': 0.9859093278646469, 'clip_ratio/low_mean': 0.06148950941860676, 'clip_ratio/low_min': 0.06148950941860676, 'clip_ratio/high_mean': 0.04861742490902543, 'clip_ratio/high_max': 0.04861742490902543, 'clip_ratio/region_mean': 0.11010693432763219, 'reward_total_mean': 0.3364430069923401, 'reward_meter_mean': 0.7479416728019714, 'reward_meter_std': 0.4325871467590332, '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_lexical_plausibility_mean': 0.904184877872467, 'reward_lexical_plausibility_std': 0.13383860886096954, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.35500001907348633, 'reward_judge_quality_std': 0.36245197057724, 'reward_total_composite_mean': 0.3364430069923401, 'reward_total_composite_std': 0.3685567080974579, 'epoch': 0.02}

 17%|█▋        | 552/3300 [2:11:10<35:18:07, 46.25s/it]INFO 04-12 12:57:22 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 553/3300 [2:11:19<26:56:39, 35.31s/it]2026-04-12 12:57:32,070 | INFO | train_grpo_train | metrics_logged mode=train step=553

                                                       
{'loss': 0.015, 'grad_norm': 10.721797943115234, 'learning_rate': 8.327272727272728e-06, 'num_tokens': 1106621.0, 'completions/mean_length': 92.25, 'completions/min_length': 83.0, 'completions/max_length': 102.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 92.25, 'completions/min_terminated_length': 83.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.7491284608840942, 'rewards/meter/std': 0.38307100534439087, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5425000190734863, 'rewards/judge_quality/std': 0.23705033957958221, 'rewards/repeat_penalty/mean': 0.7934159636497498, 'rewards/repeat_penalty/std': 0.2530416250228882, 'rewards/near_duplicate_penalty/mean': 0.9466469287872314, 'rewards/near_duplicate_penalty/std': 0.047068409621715546, 'rewards/opening_diversity/mean': 0.9609375, 'rewards/opening_diversity/std': 0.08799287676811218, 'rewards/distinct_2/mean': 0.8599124550819397, 'rewards/distinct_2/std': 0.18166913092136383, 'rewards/total_composite/mean': 0.383758544921875, 'rewards/total_composite/std': 0.2166719138622284, 'reward': 0.383758544921875, 'reward_std': 0.2166719138622284, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1297183632850647, 'sampling/sampling_logp_difference/max': 2.2496402263641357, 'sampling/importance_sampling_ratio/min': 0.10543715208768845, 'sampling/importance_sampling_ratio/mean': 1.00994873046875, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9048366099596024, 'clip_ratio/low_mean': 0.041222899220883846, 'clip_ratio/low_min': 0.041222899220883846, 'clip_ratio/high_mean': 0.08313542697578669, 'clip_ratio/high_max': 0.08313542697578669, 'clip_ratio/region_mean': 0.12435832619667053, 'reward_total_mean': 0.383758544921875, 'reward_meter_mean': 0.7491284608840942, 'reward_meter_std': 0.38307100534439087, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7934159636497498, 'reward_repeat_penalty_std': 0.2530416250228882, 'reward_near_duplicate_penalty_mean': 0.9466469287872314, 'reward_near_duplicate_penalty_std': 0.047068409621715546, 'reward_opening_diversity_mean': 0.9609375, 'reward_opening_diversity_std': 0.08799287676811218, 'reward_distinct_2_mean': 0.8599124550819397, 'reward_distinct_2_std': 0.18166913092136383, 'reward_judge_quality_mean': 0.5425000190734863, 'reward_judge_quality_std': 0.23705033957958221, 'reward_total_composite_mean': 0.383758544921875, 'reward_total_composite_std': 0.2166719138622284, 'epoch': 0.02}

 17%|█▋        | 553/3300 [2:11:20<26:56:39, 35.31s/it]INFO 04-12 12:57:32 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 554/3300 [2:11:34<22:04:50, 28.95s/it]2026-04-12 12:57:46,164 | INFO | train_grpo_train | metrics_logged mode=train step=554

                                                       
{'loss': -0.0139, 'grad_norm': 6.465330123901367, 'learning_rate': 8.324242424242425e-06, 'num_tokens': 1108361.0, 'completions/mean_length': 103.5, 'completions/min_length': 41.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 45.142860412597656, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.5443894267082214, 'rewards/meter/std': 0.4058128297328949, '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/lexical_plausibility/mean': 0.9611461162567139, 'rewards/lexical_plausibility/std': 0.1098954826593399, 'rewards/judge_quality/mean': 0.4674999713897705, 'rewards/judge_quality/std': 0.3022179901599884, 'rewards/repeat_penalty/mean': 0.9838154315948486, 'rewards/repeat_penalty/std': 0.029549043625593185, 'rewards/near_duplicate_penalty/mean': 0.9838154315948486, 'rewards/near_duplicate_penalty/std': 0.029549043625593185, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.31255239248275757, 'rewards/total_composite/std': 0.32128313183784485, 'reward': 0.31255239248275757, 'reward_std': 0.32128313183784485, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1777176856994629, 'sampling/sampling_logp_difference/max': 1.831413745880127, 'sampling/importance_sampling_ratio/min': 0.16018694639205933, 'sampling/importance_sampling_ratio/mean': 0.9967982769012451, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.195886306464672, 'clip_ratio/low_mean': 0.04569301754236221, 'clip_ratio/low_min': 0.04569301754236221, 'clip_ratio/high_mean': 0.08483864553272724, 'clip_ratio/high_max': 0.08483864553272724, 'clip_ratio/region_mean': 0.13053166307508945, 'reward_total_mean': 0.31255239248275757, 'reward_meter_mean': 0.5443894267082214, 'reward_meter_std': 0.4058128297328949, '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_lexical_plausibility_mean': 0.9611461162567139, 'reward_lexical_plausibility_std': 0.1098954826593399, 'reward_repeat_penalty_mean': 0.9838154315948486, 'reward_repeat_penalty_std': 0.029549043625593185, 'reward_near_duplicate_penalty_mean': 0.9838154315948486, 'reward_near_duplicate_penalty_std': 0.029549043625593185, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4674999713897705, 'reward_judge_quality_std': 0.3022179901599884, 'reward_total_composite_mean': 0.31255239248275757, 'reward_total_composite_std': 0.32128313183784485, 'epoch': 0.02}

 17%|█▋        | 554/3300 [2:11:34<22:04:50, 28.95s/it]INFO 04-12 12:57:46 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 555/3300 [2:11:47<18:30:13, 24.27s/it]2026-04-12 12:57:59,509 | INFO | train_grpo_train | metrics_logged mode=train step=555

                                                       
{'loss': -0.0185, 'grad_norm': 2.847667694091797, 'learning_rate': 8.321212121212123e-06, 'num_tokens': 1109962.0, 'completions/mean_length': 156.125, 'completions/min_length': 32.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 37.5, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.4670912027359009, 'rewards/meter/std': 0.38299253582954407, '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/lexical_plausibility/mean': 0.9493948221206665, 'rewards/lexical_plausibility/std': 0.08566834032535553, 'rewards/judge_quality/mean': 0.35500001907348633, 'rewards/judge_quality/std': 0.3702508807182312, 'rewards/repeat_penalty/mean': 0.9581513404846191, 'rewards/repeat_penalty/std': 0.08587849885225296, 'rewards/near_duplicate_penalty/mean': 0.9894013404846191, 'rewards/near_duplicate_penalty/std': 0.017879048362374306, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.27249038219451904, 'rewards/total_composite/std': 0.3778645992279053, 'reward': 0.27249038219451904, 'reward_std': 0.3778645992279053, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19672834873199463, 'sampling/sampling_logp_difference/max': 1.9538183212280273, 'sampling/importance_sampling_ratio/min': 0.141731858253479, 'sampling/importance_sampling_ratio/mean': 1.0331724882125854, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.115730993449688, 'clip_ratio/low_mean': 0.06789673119783401, 'clip_ratio/low_min': 0.06789673119783401, 'clip_ratio/high_mean': 0.05420168302953243, 'clip_ratio/high_max': 0.05420168302953243, 'clip_ratio/region_mean': 0.12209841422736645, 'reward_total_mean': 0.27249038219451904, 'reward_meter_mean': 0.4670912027359009, 'reward_meter_std': 0.38299253582954407, '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_lexical_plausibility_mean': 0.9493948221206665, 'reward_lexical_plausibility_std': 0.08566834032535553, 'reward_repeat_penalty_mean': 0.9581513404846191, 'reward_repeat_penalty_std': 0.08587849885225296, 'reward_near_duplicate_penalty_mean': 0.9894013404846191, 'reward_near_duplicate_penalty_std': 0.017879048362374306, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.35500001907348633, 'reward_judge_quality_std': 0.3702508807182312, 'reward_total_composite_mean': 0.27249038219451904, 'reward_total_composite_std': 0.3778645992279053, 'epoch': 0.02}

 17%|█▋        | 555/3300 [2:11:47<18:30:13, 24.27s/it]INFO 04-12 12:57:59 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 556/3300 [2:12:02<16:19:15, 21.41s/it]2026-04-12 12:58:14,263 | INFO | train_grpo_train | metrics_logged mode=train step=556

                                                       
{'loss': -0.0601, 'grad_norm': 5.946819305419922, 'learning_rate': 8.318181818181818e-06, 'num_tokens': 1111609.0, 'completions/mean_length': 112.875, 'completions/min_length': 53.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 55.857147216796875, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.7880343198776245, 'rewards/meter/std': 0.3344355523586273, '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/lexical_plausibility/mean': 0.97159743309021, 'rewards/lexical_plausibility/std': 0.08033450692892075, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.2121320366859436, 'rewards/repeat_penalty/mean': 0.9598750472068787, 'rewards/repeat_penalty/std': 0.060713496059179306, 'rewards/near_duplicate_penalty/mean': 0.9663456082344055, 'rewards/near_duplicate_penalty/std': 0.04469338804483414, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9926035404205322, 'rewards/distinct_2/std': 0.020920326933264732, 'rewards/total_composite/mean': 0.29991233348846436, 'rewards/total_composite/std': 0.2478688359260559, 'reward': 0.29991233348846436, 'reward_std': 0.24786882102489471, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15547391772270203, 'sampling/sampling_logp_difference/max': 4.527191162109375, 'sampling/importance_sampling_ratio/min': 0.01081099919974804, 'sampling/importance_sampling_ratio/mean': 1.015523910522461, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0464322119951248, 'clip_ratio/low_mean': 0.04002866800874472, 'clip_ratio/low_min': 0.04002866800874472, 'clip_ratio/high_mean': 0.07212395314127207, 'clip_ratio/high_max': 0.07212395314127207, 'clip_ratio/region_mean': 0.11215262115001678, 'reward_total_mean': 0.29991233348846436, 'reward_meter_mean': 0.7880343198776245, 'reward_meter_std': 0.3344355523586273, '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_lexical_plausibility_mean': 0.97159743309021, 'reward_lexical_plausibility_std': 0.08033450692892075, 'reward_repeat_penalty_mean': 0.9598750472068787, 'reward_repeat_penalty_std': 0.060713496059179306, 'reward_near_duplicate_penalty_mean': 0.9663456082344055, 'reward_near_duplicate_penalty_std': 0.04469338804483414, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9926035404205322, 'reward_distinct_2_std': 0.020920326933264732, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.2121320366859436, 'reward_total_composite_mean': 0.29991233348846436, 'reward_total_composite_std': 0.2478688359260559, 'epoch': 0.02}

 17%|█▋        | 556/3300 [2:12:02<16:19:15, 21.41s/it]INFO 04-12 12:58:14 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 557/3300 [2:12:16<14:39:35, 19.24s/it]2026-04-12 12:58:28,433 | INFO | train_grpo_train | metrics_logged mode=train step=557

                                                       
{'loss': -0.1467, 'grad_norm': 3.1522791385650635, 'learning_rate': 8.315151515151516e-06, 'num_tokens': 1114457.0, 'completions/mean_length': 287.0, 'completions/min_length': 191.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 212.0, 'completions/min_terminated_length': 191.0, 'completions/max_terminated_length': 220.0, 'rewards/meter/mean': 0.8281444907188416, 'rewards/meter/std': 0.3313653767108917, 'rewards/count_adherence/mean': 0.78125, 'rewards/count_adherence/std': 0.23857837915420532, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9543272256851196, 'rewards/lexical_plausibility/std': 0.12918205559253693, 'rewards/judge_quality/mean': 0.16249999403953552, 'rewards/judge_quality/std': 0.08345229923725128, 'rewards/repeat_penalty/mean': 0.2222694456577301, 'rewards/repeat_penalty/std': 0.3289700150489807, 'rewards/near_duplicate_penalty/mean': 0.7726292014122009, 'rewards/near_duplicate_penalty/std': 0.1233188584446907, 'rewards/opening_diversity/mean': 0.7336647510528564, 'rewards/opening_diversity/std': 0.29618027806282043, 'rewards/distinct_2/mean': 0.2857621908187866, 'rewards/distinct_2/std': 0.33399727940559387, 'rewards/total_composite/mean': 0.13144192099571228, 'rewards/total_composite/std': 0.09727711230516434, 'reward': 0.13144192099571228, 'reward_std': 0.09727711230516434, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07713959366083145, 'sampling/sampling_logp_difference/max': 1.5861306190490723, 'sampling/importance_sampling_ratio/min': 0.20471620559692383, 'sampling/importance_sampling_ratio/mean': 1.0107210874557495, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3740359880030155, 'clip_ratio/low_mean': 0.03940075542777777, 'clip_ratio/low_min': 0.03940075542777777, 'clip_ratio/high_mean': 0.007198953069746494, 'clip_ratio/high_max': 0.007198953069746494, 'clip_ratio/region_mean': 0.04659970849752426, 'reward_total_mean': 0.13144192099571228, 'reward_meter_mean': 0.8281444907188416, 'reward_meter_std': 0.3313653767108917, 'reward_count_adherence_mean': 0.78125, 'reward_count_adherence_std': 0.23857837915420532, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9543272256851196, 'reward_lexical_plausibility_std': 0.12918205559253693, 'reward_repeat_penalty_mean': 0.2222694456577301, 'reward_repeat_penalty_std': 0.3289700150489807, 'reward_near_duplicate_penalty_mean': 0.7726292014122009, 'reward_near_duplicate_penalty_std': 0.1233188584446907, 'reward_opening_diversity_mean': 0.7336647510528564, 'reward_opening_diversity_std': 0.29618027806282043, 'reward_distinct_2_mean': 0.2857621908187866, 'reward_distinct_2_std': 0.33399727940559387, 'reward_judge_quality_mean': 0.16249999403953552, 'reward_judge_quality_std': 0.08345229923725128, 'reward_total_composite_mean': 0.13144192099571228, 'reward_total_composite_std': 0.09727711230516434, 'epoch': 0.02}

 17%|█▋        | 557/3300 [2:12:16<14:39:35, 19.24s/it]INFO 04-12 12:58:28 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 558/3300 [2:12:31<13:45:03, 18.05s/it]2026-04-12 12:58:43,717 | INFO | train_grpo_train | metrics_logged mode=train step=558

                                                       
{'loss': -0.0201, 'grad_norm': 2.0149505138397217, 'learning_rate': 8.312121212121213e-06, 'num_tokens': 1115897.0, 'completions/mean_length': 284.0, 'completions/min_length': 35.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 56.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 111.0, 'rewards/meter/mean': 0.1431046426296234, 'rewards/meter/std': 0.1524447649717331, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 0.375, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.81473308801651, 'rewards/lexical_plausibility/std': 0.1614583134651184, 'rewards/judge_quality/mean': 0.32124999165534973, 'rewards/judge_quality/std': 0.39248064160346985, 'rewards/repeat_penalty/mean': 0.9664043188095093, 'rewards/repeat_penalty/std': 0.08752123266458511, 'rewards/near_duplicate_penalty/mean': 0.998210072517395, 'rewards/near_duplicate_penalty/std': 0.0038493694737553596, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9994441866874695, 'rewards/distinct_2/std': 0.0015720564406365156, 'rewards/total_composite/mean': 0.03383456915616989, 'rewards/total_composite/std': 0.055616602301597595, 'reward': 0.03383456915616989, 'reward_std': 0.055616602301597595, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19366790354251862, 'sampling/sampling_logp_difference/max': 1.7451801300048828, 'sampling/importance_sampling_ratio/min': 0.17461352050304413, 'sampling/importance_sampling_ratio/mean': 1.0661776065826416, 'sampling/importance_sampling_ratio/max': 1.874167799949646, 'entropy': 1.338733822107315, 'clip_ratio/low_mean': 0.018018018454313278, 'clip_ratio/low_min': 0.018018018454313278, 'clip_ratio/high_mean': 0.03905075276270509, 'clip_ratio/high_max': 0.03905075276270509, 'clip_ratio/region_mean': 0.057068771217018366, 'reward_total_mean': 0.03383456915616989, 'reward_meter_mean': 0.1431046426296234, 'reward_meter_std': 0.1524447649717331, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 0.375, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.81473308801651, 'reward_lexical_plausibility_std': 0.1614583134651184, 'reward_repeat_penalty_mean': 0.9664043188095093, 'reward_repeat_penalty_std': 0.08752123266458511, 'reward_near_duplicate_penalty_mean': 0.998210072517395, 'reward_near_duplicate_penalty_std': 0.0038493694737553596, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9994441866874695, 'reward_distinct_2_std': 0.0015720564406365156, 'reward_judge_quality_mean': 0.32124999165534973, 'reward_judge_quality_std': 0.39248064160346985, 'reward_total_composite_mean': 0.03383456915616989, 'reward_total_composite_std': 0.055616602301597595, 'epoch': 0.02}

 17%|█▋        | 558/3300 [2:12:31<13:45:03, 18.05s/it]INFO 04-12 12:58:43 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 559/3300 [2:12:45<12:47:25, 16.80s/it]2026-04-12 12:58:57,590 | INFO | train_grpo_train | metrics_logged mode=train step=559

                                                       
{'loss': 0.0633, 'grad_norm': 7.030700206756592, 'learning_rate': 8.30909090909091e-06, 'num_tokens': 1119461.0, 'completions/mean_length': 227.5, 'completions/min_length': 193.0, 'completions/max_length': 252.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 227.5, 'completions/min_terminated_length': 193.0, 'completions/max_terminated_length': 252.0, 'rewards/meter/mean': 0.7596148252487183, 'rewards/meter/std': 0.25658178329467773, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.0578637570142746, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.05345224589109421, 'rewards/repeat_penalty/mean': 0.5238463282585144, 'rewards/repeat_penalty/std': 0.29744085669517517, 'rewards/near_duplicate_penalty/mean': 0.8949508666992188, 'rewards/near_duplicate_penalty/std': 0.08561544865369797, 'rewards/opening_diversity/mean': 0.9479166865348816, 'rewards/opening_diversity/std': 0.11732383072376251, 'rewards/distinct_2/mean': 0.6009258031845093, 'rewards/distinct_2/std': 0.2808225154876709, 'rewards/total_composite/mean': 0.2752179503440857, 'rewards/total_composite/std': 0.10091891139745712, 'reward': 0.2752179503440857, 'reward_std': 0.10091891139745712, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1394757479429245, 'sampling/sampling_logp_difference/max': 3.015350580215454, 'sampling/importance_sampling_ratio/min': 0.04902864247560501, 'sampling/importance_sampling_ratio/mean': 1.0008983612060547, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7444732151925564, 'clip_ratio/low_mean': 0.04241620097309351, 'clip_ratio/low_min': 0.04241620097309351, 'clip_ratio/high_mean': 0.08743564877659082, 'clip_ratio/high_max': 0.08743564877659082, 'clip_ratio/region_mean': 0.12985184974968433, 'reward_total_mean': 0.2752179503440857, 'reward_meter_mean': 0.7596148252487183, 'reward_meter_std': 0.25658178329467773, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.0578637570142746, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.5238463282585144, 'reward_repeat_penalty_std': 0.29744085669517517, 'reward_near_duplicate_penalty_mean': 0.8949508666992188, 'reward_near_duplicate_penalty_std': 0.08561544865369797, 'reward_opening_diversity_mean': 0.9479166865348816, 'reward_opening_diversity_std': 0.11732383072376251, 'reward_distinct_2_mean': 0.6009258031845093, 'reward_distinct_2_std': 0.2808225154876709, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.05345224589109421, 'reward_total_composite_mean': 0.2752179503440857, 'reward_total_composite_std': 0.10091891139745712, 'epoch': 0.02}

 17%|█▋        | 559/3300 [2:12:45<12:47:25, 16.80s/it]INFO 04-12 12:58:57 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 560/3300 [2:13:00<12:23:18, 16.28s/it]2026-04-12 12:59:12,649 | INFO | train_grpo_train | metrics_logged mode=train step=560

                                                       
{'loss': -0.0894, 'grad_norm': 4.5670037269592285, 'learning_rate': 8.306060606060606e-06, 'num_tokens': 1122889.0, 'completions/mean_length': 279.5, 'completions/min_length': 216.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 246.2857208251953, 'completions/min_terminated_length': 216.0, 'completions/max_terminated_length': 283.0, 'rewards/meter/mean': 0.7371165752410889, 'rewards/meter/std': 0.4074375331401825, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3110969364643097, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.961893618106842, 'rewards/lexical_plausibility/std': 0.10778112709522247, 'rewards/judge_quality/mean': 0.39375001192092896, 'rewards/judge_quality/std': 0.20604698359966278, 'rewards/repeat_penalty/mean': 0.7667455673217773, 'rewards/repeat_penalty/std': 0.1332370489835739, 'rewards/near_duplicate_penalty/mean': 0.9596879482269287, 'rewards/near_duplicate_penalty/std': 0.022995855659246445, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.8332638144493103, 'rewards/distinct_2/std': 0.13802498579025269, 'rewards/total_composite/mean': 0.3322353959083557, 'rewards/total_composite/std': 0.2467871457338333, 'reward': 0.3322353959083557, 'reward_std': 0.2467871606349945, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15317615866661072, 'sampling/sampling_logp_difference/max': 2.7960309982299805, 'sampling/importance_sampling_ratio/min': 0.17604616284370422, 'sampling/importance_sampling_ratio/mean': 1.0157796144485474, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8752892836928368, 'clip_ratio/low_mean': 0.03733076713979244, 'clip_ratio/low_min': 0.03733076713979244, 'clip_ratio/high_mean': 0.08274878561496735, 'clip_ratio/high_max': 0.08274878561496735, 'clip_ratio/region_mean': 0.12007955275475979, 'reward_total_mean': 0.3322353959083557, 'reward_meter_mean': 0.7371165752410889, 'reward_meter_std': 0.4074375331401825, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3110969364643097, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.961893618106842, 'reward_lexical_plausibility_std': 0.10778112709522247, 'reward_repeat_penalty_mean': 0.7667455673217773, 'reward_repeat_penalty_std': 0.1332370489835739, 'reward_near_duplicate_penalty_mean': 0.9596879482269287, 'reward_near_duplicate_penalty_std': 0.022995855659246445, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.8332638144493103, 'reward_distinct_2_std': 0.13802498579025269, 'reward_judge_quality_mean': 0.39375001192092896, 'reward_judge_quality_std': 0.20604698359966278, 'reward_total_composite_mean': 0.3322353959083557, 'reward_total_composite_std': 0.2467871457338333, 'epoch': 0.02}

 17%|█▋        | 560/3300 [2:13:00<12:23:18, 16.28s/it]INFO 04-12 12:59:12 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 561/3300 [2:13:10<10:54:04, 14.33s/it]2026-04-12 12:59:22,428 | INFO | train_grpo_train | metrics_logged mode=train step=561

                                                       
{'loss': 0.0637, 'grad_norm': 15.194409370422363, 'learning_rate': 8.303030303030305e-06, 'num_tokens': 1124482.0, 'completions/mean_length': 46.125, 'completions/min_length': 40.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.125, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.859315037727356, 'rewards/meter/std': 0.23800811171531677, '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/lexical_plausibility/mean': 0.9320663809776306, 'rewards/lexical_plausibility/std': 0.16831834614276886, 'rewards/judge_quality/mean': 0.7350000143051147, 'rewards/judge_quality/std': 0.231084406375885, 'rewards/repeat_penalty/mean': 0.9562109708786011, 'rewards/repeat_penalty/std': 0.05390492081642151, 'rewards/near_duplicate_penalty/mean': 0.9744315147399902, 'rewards/near_duplicate_penalty/std': 0.02383003756403923, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9807209968566895, 'rewards/distinct_2/std': 0.0357571542263031, 'rewards/total_composite/mean': 0.5776807069778442, 'rewards/total_composite/std': 0.3228343427181244, 'reward': 0.5776807069778442, 'reward_std': 0.322834312915802, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17364515364170074, 'sampling/sampling_logp_difference/max': 2.161709785461426, 'sampling/importance_sampling_ratio/min': 0.11512810736894608, 'sampling/importance_sampling_ratio/mean': 0.9876819849014282, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0650817900896072, 'clip_ratio/low_mean': 0.061446020379662514, 'clip_ratio/low_min': 0.061446020379662514, 'clip_ratio/high_mean': 0.11508160457015038, 'clip_ratio/high_max': 0.11508160457015038, 'clip_ratio/region_mean': 0.1765276249498129, 'reward_total_mean': 0.5776807069778442, 'reward_meter_mean': 0.859315037727356, 'reward_meter_std': 0.23800811171531677, '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_lexical_plausibility_mean': 0.9320663809776306, 'reward_lexical_plausibility_std': 0.16831834614276886, 'reward_repeat_penalty_mean': 0.9562109708786011, 'reward_repeat_penalty_std': 0.05390492081642151, 'reward_near_duplicate_penalty_mean': 0.9744315147399902, 'reward_near_duplicate_penalty_std': 0.02383003756403923, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9807209968566895, 'reward_distinct_2_std': 0.0357571542263031, 'reward_judge_quality_mean': 0.7350000143051147, 'reward_judge_quality_std': 0.231084406375885, 'reward_total_composite_mean': 0.5776807069778442, 'reward_total_composite_std': 0.3228343427181244, 'epoch': 0.02}

 17%|█▋        | 561/3300 [2:13:10<10:54:04, 14.33s/it]INFO 04-12 12:59:22 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 562/3300 [2:13:21<10:15:54, 13.50s/it]2026-04-12 12:59:33,989 | INFO | train_grpo_train | metrics_logged mode=train step=562

                                                       
{'loss': 0.0585, 'grad_norm': 12.776339530944824, 'learning_rate': 8.3e-06, 'num_tokens': 1126206.0, 'completions/mean_length': 52.5, 'completions/min_length': 47.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.5, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.6997398138046265, 'rewards/meter/std': 0.3450835347175598, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6637499928474426, 'rewards/judge_quality/std': 0.28465205430984497, 'rewards/repeat_penalty/mean': 0.966732919216156, 'rewards/repeat_penalty/std': 0.04837942123413086, 'rewards/near_duplicate_penalty/mean': 0.9843859076499939, 'rewards/near_duplicate_penalty/std': 0.02186666615307331, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9816849827766418, 'rewards/distinct_2/std': 0.03391282260417938, 'rewards/total_composite/mean': 0.4740862548351288, 'rewards/total_composite/std': 0.33488765358924866, 'reward': 0.4740862548351288, 'reward_std': 0.33488768339157104, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16140180826187134, 'sampling/sampling_logp_difference/max': 1.4580715894699097, 'sampling/importance_sampling_ratio/min': 0.23268455266952515, 'sampling/importance_sampling_ratio/mean': 1.0163640975952148, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1352339908480644, 'clip_ratio/low_mean': 0.06900547398254275, 'clip_ratio/low_min': 0.06900547398254275, 'clip_ratio/high_mean': 0.05386254005134106, 'clip_ratio/high_max': 0.05386254005134106, 'clip_ratio/region_mean': 0.12286801403388381, 'reward_total_mean': 0.4740862548351288, 'reward_meter_mean': 0.6997398138046265, 'reward_meter_std': 0.3450835347175598, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.966732919216156, 'reward_repeat_penalty_std': 0.04837942123413086, 'reward_near_duplicate_penalty_mean': 0.9843859076499939, 'reward_near_duplicate_penalty_std': 0.02186666615307331, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9816849827766418, 'reward_distinct_2_std': 0.03391282260417938, 'reward_judge_quality_mean': 0.6637499928474426, 'reward_judge_quality_std': 0.28465205430984497, 'reward_total_composite_mean': 0.4740862548351288, 'reward_total_composite_std': 0.33488765358924866, 'epoch': 0.02}

 17%|█▋        | 562/3300 [2:13:21<10:15:54, 13.50s/it]INFO 04-12 12:59:34 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 563/3300 [2:13:32<9:38:47, 12.69s/it] 2026-04-12 12:59:44,792 | INFO | train_grpo_train | metrics_logged mode=train step=563

                                                      
{'loss': 0.0082, 'grad_norm': 17.09084701538086, 'learning_rate': 8.296969696969697e-06, 'num_tokens': 1128294.0, 'completions/mean_length': 87.0, 'completions/min_length': 74.0, 'completions/max_length': 104.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 87.0, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 104.0, 'rewards/meter/mean': 0.394396036863327, 'rewards/meter/std': 0.2646758556365967, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.13363061845302582, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.0353553369641304, 'rewards/repeat_penalty/mean': 0.6897436380386353, 'rewards/repeat_penalty/std': 0.2753545045852661, 'rewards/near_duplicate_penalty/mean': 0.903167188167572, 'rewards/near_duplicate_penalty/std': 0.08225918561220169, 'rewards/opening_diversity/mean': 0.9609375, 'rewards/opening_diversity/std': 0.08799287676811218, 'rewards/distinct_2/mean': 0.7783054113388062, 'rewards/distinct_2/std': 0.20354458689689636, 'rewards/total_composite/mean': 0.12317164987325668, 'rewards/total_composite/std': 0.08558519929647446, 'reward': 0.12317164987325668, 'reward_std': 0.08558519929647446, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16062937676906586, 'sampling/sampling_logp_difference/max': 2.943558692932129, 'sampling/importance_sampling_ratio/min': 0.052677929401397705, 'sampling/importance_sampling_ratio/mean': 1.0046311616897583, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6217370927333832, 'clip_ratio/low_mean': 0.09821091499179602, 'clip_ratio/low_min': 0.09821091499179602, 'clip_ratio/high_mean': 0.04175998363643885, 'clip_ratio/high_max': 0.04175998363643885, 'clip_ratio/region_mean': 0.13997089862823486, 'reward_total_mean': 0.12317164987325668, 'reward_meter_mean': 0.394396036863327, 'reward_meter_std': 0.2646758556365967, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.13363061845302582, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6897436380386353, 'reward_repeat_penalty_std': 0.2753545045852661, 'reward_near_duplicate_penalty_mean': 0.903167188167572, 'reward_near_duplicate_penalty_std': 0.08225918561220169, 'reward_opening_diversity_mean': 0.9609375, 'reward_opening_diversity_std': 0.08799287676811218, 'reward_distinct_2_mean': 0.7783054113388062, 'reward_distinct_2_std': 0.20354458689689636, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.0353553369641304, 'reward_total_composite_mean': 0.12317164987325668, 'reward_total_composite_std': 0.08558519929647446, 'epoch': 0.02}

 17%|█▋        | 563/3300 [2:13:32<9:38:47, 12.69s/it]INFO 04-12 12:59:45 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 564/3300 [2:13:48<10:19:41, 13.59s/it]2026-04-12 13:00:00,492 | INFO | train_grpo_train | metrics_logged mode=train step=564

                                                       
{'loss': -0.0646, 'grad_norm': 3.1676385402679443, 'learning_rate': 8.293939393939395e-06, 'num_tokens': 1129835.0, 'completions/mean_length': 161.625, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 44.833335876464844, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.36686593294143677, 'rewards/meter/std': 0.36617180705070496, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.944107711315155, 'rewards/lexical_plausibility/std': 0.10829658061265945, 'rewards/judge_quality/mean': 0.26500001549720764, 'rewards/judge_quality/std': 0.29267242550849915, 'rewards/repeat_penalty/mean': 0.9262036085128784, 'rewards/repeat_penalty/std': 0.11174186319112778, 'rewards/near_duplicate_penalty/mean': 0.9790786504745483, 'rewards/near_duplicate_penalty/std': 0.03449733927845955, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9763580560684204, 'rewards/distinct_2/std': 0.06364420056343079, 'rewards/total_composite/mean': 0.07410462200641632, 'rewards/total_composite/std': 0.07805755734443665, 'reward': 0.07410462200641632, 'reward_std': 0.07805755734443665, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14709855616092682, 'sampling/sampling_logp_difference/max': 1.3158307075500488, 'sampling/importance_sampling_ratio/min': 0.2682513892650604, 'sampling/importance_sampling_ratio/mean': 1.0264726877212524, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0641838908195496, 'clip_ratio/low_mean': 0.05612799432128668, 'clip_ratio/low_min': 0.05612799432128668, 'clip_ratio/high_mean': 0.06306612212210894, 'clip_ratio/high_max': 0.06306612212210894, 'clip_ratio/region_mean': 0.11919411644339561, 'reward_total_mean': 0.07410462200641632, 'reward_meter_mean': 0.36686593294143677, 'reward_meter_std': 0.36617180705070496, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.944107711315155, 'reward_lexical_plausibility_std': 0.10829658061265945, 'reward_repeat_penalty_mean': 0.9262036085128784, 'reward_repeat_penalty_std': 0.11174186319112778, 'reward_near_duplicate_penalty_mean': 0.9790786504745483, 'reward_near_duplicate_penalty_std': 0.03449733927845955, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9763580560684204, 'reward_distinct_2_std': 0.06364420056343079, 'reward_judge_quality_mean': 0.26500001549720764, 'reward_judge_quality_std': 0.29267242550849915, 'reward_total_composite_mean': 0.07410462200641632, 'reward_total_composite_std': 0.07805755734443665, 'epoch': 0.02}

 17%|█▋        | 564/3300 [2:13:48<10:19:41, 13.59s/it]INFO 04-12 13:00:00 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 565/3300 [2:14:01<10:12:10, 13.43s/it]2026-04-12 13:00:13,542 | INFO | train_grpo_train | metrics_logged mode=train step=565

                                                       
{'loss': 0.1062, 'grad_norm': 5.6130051612854, 'learning_rate': 8.290909090909092e-06, 'num_tokens': 1133639.0, 'completions/mean_length': 276.5, 'completions/min_length': 237.0, 'completions/max_length': 343.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 276.5, 'completions/min_terminated_length': 237.0, 'completions/max_terminated_length': 343.0, 'rewards/meter/mean': 0.7869507074356079, 'rewards/meter/std': 0.14151926338672638, 'rewards/count_adherence/mean': 0.8541666865348816, 'rewards/count_adherence/std': 0.13908715546131134, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.547345757484436, 'rewards/repeat_penalty/std': 0.2718929946422577, 'rewards/near_duplicate_penalty/mean': 0.937583327293396, 'rewards/near_duplicate_penalty/std': 0.05010360851883888, 'rewards/opening_diversity/mean': 0.998046875, 'rewards/opening_diversity/std': 0.005524271633476019, 'rewards/distinct_2/mean': 0.6103899478912354, 'rewards/distinct_2/std': 0.265046626329422, 'rewards/total_composite/mean': 0.25660911202430725, 'rewards/total_composite/std': 0.08675975352525711, 'reward': 0.25660911202430725, 'reward_std': 0.08675975352525711, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14434683322906494, 'sampling/sampling_logp_difference/max': 3.466536521911621, 'sampling/importance_sampling_ratio/min': 0.031224992126226425, 'sampling/importance_sampling_ratio/mean': 1.0100514888763428, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7649172767996788, 'clip_ratio/low_mean': 0.06368709448724985, 'clip_ratio/low_min': 0.06368709448724985, 'clip_ratio/high_mean': 0.07652063854038715, 'clip_ratio/high_max': 0.07652063854038715, 'clip_ratio/region_mean': 0.140207733027637, 'reward_total_mean': 0.25660911202430725, 'reward_meter_mean': 0.7869507074356079, 'reward_meter_std': 0.14151926338672638, 'reward_count_adherence_mean': 0.8541666865348816, 'reward_count_adherence_std': 0.13908715546131134, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.547345757484436, 'reward_repeat_penalty_std': 0.2718929946422577, 'reward_near_duplicate_penalty_mean': 0.937583327293396, 'reward_near_duplicate_penalty_std': 0.05010360851883888, 'reward_opening_diversity_mean': 0.998046875, 'reward_opening_diversity_std': 0.005524271633476019, 'reward_distinct_2_mean': 0.6103899478912354, 'reward_distinct_2_std': 0.265046626329422, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.25660911202430725, 'reward_total_composite_std': 0.08675975352525711, 'epoch': 0.02}

 17%|█▋        | 565/3300 [2:14:01<10:12:10, 13.43s/it]INFO 04-12 13:00:13 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 566/3300 [2:14:15<10:23:25, 13.68s/it]2026-04-12 13:00:27,808 | INFO | train_grpo_train | metrics_logged mode=train step=566

                                                       
{'loss': -0.064, 'grad_norm': 2.637990713119507, 'learning_rate': 8.287878787878787e-06, 'num_tokens': 1135129.0, 'completions/mean_length': 163.25, 'completions/min_length': 41.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 47.0, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.6106027364730835, 'rewards/meter/std': 0.4446389377117157, '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/lexical_plausibility/mean': 0.9282266497612, 'rewards/lexical_plausibility/std': 0.1329856663942337, 'rewards/judge_quality/mean': 0.5637500286102295, 'rewards/judge_quality/std': 0.37209591269493103, 'rewards/repeat_penalty/mean': 0.9416448473930359, 'rewards/repeat_penalty/std': 0.11396066099405289, 'rewards/near_duplicate_penalty/mean': 0.980570912361145, 'rewards/near_duplicate_penalty/std': 0.0292997807264328, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9886243939399719, 'rewards/distinct_2/std': 0.024782273918390274, 'rewards/total_composite/mean': 0.4156671464443207, 'rewards/total_composite/std': 0.3648478388786316, 'reward': 0.4156671464443207, 'reward_std': 0.3648478388786316, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1527395099401474, 'sampling/sampling_logp_difference/max': 1.5353171825408936, 'sampling/importance_sampling_ratio/min': 0.21538737416267395, 'sampling/importance_sampling_ratio/mean': 1.01309072971344, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8335897997021675, 'clip_ratio/low_mean': 0.031235750764608383, 'clip_ratio/low_min': 0.031235750764608383, 'clip_ratio/high_mean': 0.07875300478190184, 'clip_ratio/high_max': 0.07875300478190184, 'clip_ratio/region_mean': 0.10998875554651022, 'reward_total_mean': 0.4156671464443207, 'reward_meter_mean': 0.6106027364730835, 'reward_meter_std': 0.4446389377117157, '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_lexical_plausibility_mean': 0.9282266497612, 'reward_lexical_plausibility_std': 0.1329856663942337, 'reward_repeat_penalty_mean': 0.9416448473930359, 'reward_repeat_penalty_std': 0.11396066099405289, 'reward_near_duplicate_penalty_mean': 0.980570912361145, 'reward_near_duplicate_penalty_std': 0.0292997807264328, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9886243939399719, 'reward_distinct_2_std': 0.024782273918390274, 'reward_judge_quality_mean': 0.5637500286102295, 'reward_judge_quality_std': 0.37209591269493103, 'reward_total_composite_mean': 0.4156671464443207, 'reward_total_composite_std': 0.3648478388786316, 'epoch': 0.02}

 17%|█▋        | 566/3300 [2:14:15<10:23:25, 13.68s/it]INFO 04-12 13:00:28 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 567/3300 [2:14:25<9:31:19, 12.54s/it] 2026-04-12 13:00:37,693 | INFO | train_grpo_train | metrics_logged mode=train step=567

                                                      
{'loss': 0.071, 'grad_norm': 13.785236358642578, 'learning_rate': 8.284848484848486e-06, 'num_tokens': 1136758.0, 'completions/mean_length': 44.625, 'completions/min_length': 38.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.625, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.9689449071884155, 'rewards/meter/std': 0.04034477844834328, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6349999904632568, 'rewards/judge_quality/std': 0.3184336721897125, 'rewards/repeat_penalty/mean': 0.6969336867332458, 'rewards/repeat_penalty/std': 0.37751781940460205, 'rewards/near_duplicate_penalty/mean': 0.8623055219650269, 'rewards/near_duplicate_penalty/std': 0.16690663993358612, 'rewards/opening_diversity/mean': 0.859375, 'rewards/opening_diversity/std': 0.2259652018547058, 'rewards/distinct_2/mean': 0.8451367616653442, 'rewards/distinct_2/std': 0.26540008187294006, 'rewards/total_composite/mean': 0.6106916666030884, 'rewards/total_composite/std': 0.3000105917453766, 'reward': 0.6106916666030884, 'reward_std': 0.3000105917453766, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14804808795452118, 'sampling/sampling_logp_difference/max': 1.9665656089782715, 'sampling/importance_sampling_ratio/min': 0.139936625957489, 'sampling/importance_sampling_ratio/mean': 1.0200014114379883, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9125291332602501, 'clip_ratio/low_mean': 0.05262185260653496, 'clip_ratio/low_min': 0.05262185260653496, 'clip_ratio/high_mean': 0.06844110134989023, 'clip_ratio/high_max': 0.06844110134989023, 'clip_ratio/region_mean': 0.12106295395642519, 'reward_total_mean': 0.6106916666030884, 'reward_meter_mean': 0.9689449071884155, 'reward_meter_std': 0.04034477844834328, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6969336867332458, 'reward_repeat_penalty_std': 0.37751781940460205, 'reward_near_duplicate_penalty_mean': 0.8623055219650269, 'reward_near_duplicate_penalty_std': 0.16690663993358612, 'reward_opening_diversity_mean': 0.859375, 'reward_opening_diversity_std': 0.2259652018547058, 'reward_distinct_2_mean': 0.8451367616653442, 'reward_distinct_2_std': 0.26540008187294006, 'reward_judge_quality_mean': 0.6349999904632568, 'reward_judge_quality_std': 0.3184336721897125, 'reward_total_composite_mean': 0.6106916666030884, 'reward_total_composite_std': 0.3000105917453766, 'epoch': 0.02}

 17%|█▋        | 567/3300 [2:14:25<9:31:19, 12.54s/it]INFO 04-12 13:00:37 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 568/3300 [2:14:36<9:06:30, 12.00s/it]2026-04-12 13:00:48,433 | INFO | train_grpo_train | metrics_logged mode=train step=568

                                                      
{'loss': -0.0352, 'grad_norm': 7.714836597442627, 'learning_rate': 8.281818181818182e-06, 'num_tokens': 1139410.0, 'completions/mean_length': 125.5, 'completions/min_length': 105.0, 'completions/max_length': 142.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 125.5, 'completions/min_terminated_length': 105.0, 'completions/max_terminated_length': 142.0, 'rewards/meter/mean': 0.896450400352478, 'rewards/meter/std': 0.165233314037323, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4662500023841858, 'rewards/judge_quality/std': 0.1743508279323578, 'rewards/repeat_penalty/mean': 0.907419741153717, 'rewards/repeat_penalty/std': 0.18419818580150604, 'rewards/near_duplicate_penalty/mean': 0.9728611707687378, 'rewards/near_duplicate_penalty/std': 0.03465347737073898, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9366430044174194, 'rewards/distinct_2/std': 0.14353594183921814, 'rewards/total_composite/mean': 0.3970481753349304, 'rewards/total_composite/std': 0.12406831979751587, 'reward': 0.3970481753349304, 'reward_std': 0.12406831234693527, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14454703032970428, 'sampling/sampling_logp_difference/max': 2.0742719173431396, 'sampling/importance_sampling_ratio/min': 0.1256478726863861, 'sampling/importance_sampling_ratio/mean': 1.007752776145935, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0344690158963203, 'clip_ratio/low_mean': 0.06165926530957222, 'clip_ratio/low_min': 0.06165926530957222, 'clip_ratio/high_mean': 0.07605502009391785, 'clip_ratio/high_max': 0.07605502009391785, 'clip_ratio/region_mean': 0.13771428540349007, 'reward_total_mean': 0.3970481753349304, 'reward_meter_mean': 0.896450400352478, 'reward_meter_std': 0.165233314037323, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.907419741153717, 'reward_repeat_penalty_std': 0.18419818580150604, 'reward_near_duplicate_penalty_mean': 0.9728611707687378, 'reward_near_duplicate_penalty_std': 0.03465347737073898, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9366430044174194, 'reward_distinct_2_std': 0.14353594183921814, 'reward_judge_quality_mean': 0.4662500023841858, 'reward_judge_quality_std': 0.1743508279323578, 'reward_total_composite_mean': 0.3970481753349304, 'reward_total_composite_std': 0.12406831979751587, 'epoch': 0.02}

 17%|█▋        | 568/3300 [2:14:36<9:06:30, 12.00s/it]INFO 04-12 13:00:48 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 569/3300 [2:14:47<8:57:14, 11.80s/it]2026-04-12 13:00:59,773 | INFO | train_grpo_train | metrics_logged mode=train step=569

                                                      
{'loss': 0.0284, 'grad_norm': 7.0395097732543945, 'learning_rate': 8.27878787878788e-06, 'num_tokens': 1141800.0, 'completions/mean_length': 107.75, 'completions/min_length': 96.0, 'completions/max_length': 124.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 107.75, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 124.0, 'rewards/meter/mean': 0.5035848617553711, 'rewards/meter/std': 0.2208067774772644, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4749999940395355, 'rewards/judge_quality/std': 0.1752549111843109, 'rewards/repeat_penalty/mean': 0.8326623439788818, 'rewards/repeat_penalty/std': 0.3059518039226532, 'rewards/near_duplicate_penalty/mean': 0.9547679424285889, 'rewards/near_duplicate_penalty/std': 0.09025280177593231, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.8751956224441528, 'rewards/distinct_2/std': 0.22261351346969604, 'rewards/total_composite/mean': 0.20396524667739868, 'rewards/total_composite/std': 0.1100090891122818, 'reward': 0.20396524667739868, 'reward_std': 0.1100090891122818, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12819813191890717, 'sampling/sampling_logp_difference/max': 2.7097034454345703, 'sampling/importance_sampling_ratio/min': 0.06655654311180115, 'sampling/importance_sampling_ratio/mean': 0.9968386292457581, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6736838817596436, 'clip_ratio/low_mean': 0.07226160261780024, 'clip_ratio/low_min': 0.07226160261780024, 'clip_ratio/high_mean': 0.037748207338154316, 'clip_ratio/high_max': 0.037748207338154316, 'clip_ratio/region_mean': 0.11000980995595455, 'reward_total_mean': 0.20396524667739868, 'reward_meter_mean': 0.5035848617553711, 'reward_meter_std': 0.2208067774772644, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8326623439788818, 'reward_repeat_penalty_std': 0.3059518039226532, 'reward_near_duplicate_penalty_mean': 0.9547679424285889, 'reward_near_duplicate_penalty_std': 0.09025280177593231, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.8751956224441528, 'reward_distinct_2_std': 0.22261351346969604, 'reward_judge_quality_mean': 0.4749999940395355, 'reward_judge_quality_std': 0.1752549111843109, 'reward_total_composite_mean': 0.20396524667739868, 'reward_total_composite_std': 0.1100090891122818, 'epoch': 0.02}

 17%|█▋        | 569/3300 [2:14:47<8:57:14, 11.80s/it]INFO 04-12 13:01:00 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 570/3300 [2:15:00<9:13:12, 12.16s/it]2026-04-12 13:01:12,760 | INFO | train_grpo_train | metrics_logged mode=train step=570

                                                      
{'loss': -0.0682, 'grad_norm': 3.0202219486236572, 'learning_rate': 8.275757575757577e-06, 'num_tokens': 1143087.0, 'completions/mean_length': 81.875, 'completions/min_length': 18.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 20.428571701049805, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.71962970495224, 'rewards/meter/std': 0.35546237230300903, '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/lexical_plausibility/mean': 0.972630500793457, 'rewards/lexical_plausibility/std': 0.07741261273622513, 'rewards/judge_quality/mean': 0.8299999833106995, 'rewards/judge_quality/std': 0.31545883417129517, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6782770752906799, 'rewards/total_composite/std': 0.3369387090206146, 'reward': 0.6782770752906799, 'reward_std': 0.3369387090206146, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1483437567949295, 'sampling/sampling_logp_difference/max': 1.1063976287841797, 'sampling/importance_sampling_ratio/min': 0.3307483196258545, 'sampling/importance_sampling_ratio/mean': 1.0200278759002686, 'sampling/importance_sampling_ratio/max': 1.6298166513442993, 'entropy': 0.8440438508987427, 'clip_ratio/low_mean': 0.0069444444961845875, 'clip_ratio/low_min': 0.0069444444961845875, 'clip_ratio/high_mean': 0.1421447629109025, 'clip_ratio/high_max': 0.1421447629109025, 'clip_ratio/region_mean': 0.1490892074070871, 'reward_total_mean': 0.6782770752906799, 'reward_meter_mean': 0.71962970495224, 'reward_meter_std': 0.35546237230300903, '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_lexical_plausibility_mean': 0.972630500793457, 'reward_lexical_plausibility_std': 0.07741261273622513, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8299999833106995, 'reward_judge_quality_std': 0.31545883417129517, 'reward_total_composite_mean': 0.6782770752906799, 'reward_total_composite_std': 0.3369387090206146, 'epoch': 0.02}

 17%|█▋        | 570/3300 [2:15:00<9:13:12, 12.16s/it]INFO 04-12 13:01:13 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 571/3300 [2:15:15<9:53:53, 13.06s/it]2026-04-12 13:01:27,915 | INFO | train_grpo_train | metrics_logged mode=train step=571

                                                      
{'loss': -0.0724, 'grad_norm': 2.476228713989258, 'learning_rate': 8.272727272727274e-06, 'num_tokens': 1144658.0, 'completions/mean_length': 166.375, 'completions/min_length': 46.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 51.16666793823242, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.6900045871734619, 'rewards/meter/std': 0.4044930934906006, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9468242526054382, 'rewards/lexical_plausibility/std': 0.10111001133918762, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.23260942101478577, 'rewards/repeat_penalty/mean': 0.9614250659942627, 'rewards/repeat_penalty/std': 0.08640110492706299, 'rewards/near_duplicate_penalty/mean': 0.9926750659942627, 'rewards/near_duplicate_penalty/std': 0.013261236250400543, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.29301339387893677, 'rewards/total_composite/std': 0.2625063359737396, 'reward': 0.29301339387893677, 'reward_std': 0.2625063359737396, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16469305753707886, 'sampling/sampling_logp_difference/max': 2.1570959091186523, 'sampling/importance_sampling_ratio/min': 0.11566052585840225, 'sampling/importance_sampling_ratio/mean': 1.035529375076294, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.016656644642353, 'clip_ratio/low_mean': 0.040816327556967735, 'clip_ratio/low_min': 0.040816327556967735, 'clip_ratio/high_mean': 0.08386882115155458, 'clip_ratio/high_max': 0.08386882115155458, 'clip_ratio/region_mean': 0.12468514870852232, 'reward_total_mean': 0.29301339387893677, 'reward_meter_mean': 0.6900045871734619, 'reward_meter_std': 0.4044930934906006, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9468242526054382, 'reward_lexical_plausibility_std': 0.10111001133918762, 'reward_repeat_penalty_mean': 0.9614250659942627, 'reward_repeat_penalty_std': 0.08640110492706299, 'reward_near_duplicate_penalty_mean': 0.9926750659942627, 'reward_near_duplicate_penalty_std': 0.013261236250400543, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.23260942101478577, 'reward_total_composite_mean': 0.29301339387893677, 'reward_total_composite_std': 0.2625063359737396, 'epoch': 0.02}

 17%|█▋        | 571/3300 [2:15:15<9:53:53, 13.06s/it]INFO 04-12 13:01:28 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 572/3300 [2:15:28<9:44:56, 12.87s/it]2026-04-12 13:01:40,335 | INFO | train_grpo_train | metrics_logged mode=train step=572

                                                      
{'loss': -0.0547, 'grad_norm': 9.258003234863281, 'learning_rate': 8.269696969696971e-06, 'num_tokens': 1147431.0, 'completions/mean_length': 151.625, 'completions/min_length': 47.0, 'completions/max_length': 183.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 151.625, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 183.0, 'rewards/meter/mean': 0.809888482093811, 'rewards/meter/std': 0.2024429887533188, 'rewards/count_adherence/mean': 0.8928571343421936, 'rewards/count_adherence/std': 0.30304577946662903, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4874999523162842, 'rewards/judge_quality/std': 0.1060660183429718, 'rewards/repeat_penalty/mean': 0.9014173746109009, 'rewards/repeat_penalty/std': 0.06960899382829666, 'rewards/near_duplicate_penalty/mean': 0.9741859436035156, 'rewards/near_duplicate_penalty/std': 0.01830022595822811, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9247065782546997, 'rewards/distinct_2/std': 0.05930177494883537, 'rewards/total_composite/mean': 0.2677019536495209, 'rewards/total_composite/std': 0.23452313244342804, 'reward': 0.2677019536495209, 'reward_std': 0.23452313244342804, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14180618524551392, 'sampling/sampling_logp_difference/max': 2.9570159912109375, 'sampling/importance_sampling_ratio/min': 0.051973775029182434, 'sampling/importance_sampling_ratio/mean': 1.0224672555923462, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8546986281871796, 'clip_ratio/low_mean': 0.04997908137738705, 'clip_ratio/low_min': 0.04997908137738705, 'clip_ratio/high_mean': 0.07101702503859997, 'clip_ratio/high_max': 0.07101702503859997, 'clip_ratio/region_mean': 0.12099610641598701, 'reward_total_mean': 0.2677019536495209, 'reward_meter_mean': 0.809888482093811, 'reward_meter_std': 0.2024429887533188, 'reward_count_adherence_mean': 0.8928571343421936, 'reward_count_adherence_std': 0.30304577946662903, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9014173746109009, 'reward_repeat_penalty_std': 0.06960899382829666, 'reward_near_duplicate_penalty_mean': 0.9741859436035156, 'reward_near_duplicate_penalty_std': 0.01830022595822811, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9247065782546997, 'reward_distinct_2_std': 0.05930177494883537, 'reward_judge_quality_mean': 0.4874999523162842, 'reward_judge_quality_std': 0.1060660183429718, 'reward_total_composite_mean': 0.2677019536495209, 'reward_total_composite_std': 0.23452313244342804, 'epoch': 0.02}

 17%|█▋        | 572/3300 [2:15:28<9:44:56, 12.87s/it]INFO 04-12 13:01:40 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 573/3300 [2:15:41<9:56:37, 13.13s/it]2026-04-12 13:01:54,071 | INFO | train_grpo_train | metrics_logged mode=train step=573

                                                      
{'loss': -0.0367, 'grad_norm': 3.1884655952453613, 'learning_rate': 8.266666666666667e-06, 'num_tokens': 1148939.0, 'completions/mean_length': 161.5, 'completions/min_length': 42.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 44.66666793823242, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.2651941776275635, 'rewards/meter/std': 0.42671820521354675, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9270649552345276, 'rewards/lexical_plausibility/std': 0.14048178493976593, 'rewards/judge_quality/mean': 0.5637500286102295, 'rewards/judge_quality/std': 0.37209591269493103, 'rewards/repeat_penalty/mean': 0.9437205195426941, 'rewards/repeat_penalty/std': 0.08234898746013641, 'rewards/near_duplicate_penalty/mean': 0.9856542348861694, 'rewards/near_duplicate_penalty/std': 0.013923265039920807, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9893162250518799, 'rewards/distinct_2/std': 0.03021823801100254, 'rewards/total_composite/mean': 0.16601510345935822, 'rewards/total_composite/std': 0.2700558006763458, 'reward': 0.16601510345935822, 'reward_std': 0.2700558006763458, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1659148633480072, 'sampling/sampling_logp_difference/max': 1.5692558288574219, 'sampling/importance_sampling_ratio/min': 0.20820006728172302, 'sampling/importance_sampling_ratio/mean': 1.0147768259048462, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6306677088141441, 'clip_ratio/low_mean': 0.04621415585279465, 'clip_ratio/low_min': 0.04621415585279465, 'clip_ratio/high_mean': 0.040568181313574314, 'clip_ratio/high_max': 0.040568181313574314, 'clip_ratio/region_mean': 0.08678233716636896, 'reward_total_mean': 0.16601510345935822, 'reward_meter_mean': 0.2651941776275635, 'reward_meter_std': 0.42671820521354675, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9270649552345276, 'reward_lexical_plausibility_std': 0.14048178493976593, 'reward_repeat_penalty_mean': 0.9437205195426941, 'reward_repeat_penalty_std': 0.08234898746013641, 'reward_near_duplicate_penalty_mean': 0.9856542348861694, 'reward_near_duplicate_penalty_std': 0.013923265039920807, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9893162250518799, 'reward_distinct_2_std': 0.03021823801100254, 'reward_judge_quality_mean': 0.5637500286102295, 'reward_judge_quality_std': 0.37209591269493103, 'reward_total_composite_mean': 0.16601510345935822, 'reward_total_composite_std': 0.2700558006763458, 'epoch': 0.02}

 17%|█▋        | 573/3300 [2:15:42<9:56:37, 13.13s/it]INFO 04-12 13:01:54 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 574/3300 [2:15:53<9:36:36, 12.69s/it]2026-04-12 13:02:05,746 | INFO | train_grpo_train | metrics_logged mode=train step=574

                                                      
{'loss': 0.0198, 'grad_norm': 16.67266082763672, 'learning_rate': 8.263636363636366e-06, 'num_tokens': 1151467.0, 'completions/mean_length': 119.0, 'completions/min_length': 98.0, 'completions/max_length': 138.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 119.0, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 138.0, 'rewards/meter/mean': 0.7381832599639893, 'rewards/meter/std': 0.2610228657722473, 'rewards/count_adherence/mean': 0.8999999761581421, 'rewards/count_adherence/std': 0.1511857807636261, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.09258200973272324, 'rewards/repeat_penalty/mean': 0.9304074048995972, 'rewards/repeat_penalty/std': 0.13754808902740479, 'rewards/near_duplicate_penalty/mean': 0.9826213121414185, 'rewards/near_duplicate_penalty/std': 0.017901146784424782, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9549735188484192, 'rewards/distinct_2/std': 0.1023939698934555, 'rewards/total_composite/mean': 0.21185347437858582, 'rewards/total_composite/std': 0.1496955156326294, 'reward': 0.21185347437858582, 'reward_std': 0.1496955007314682, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1336987316608429, 'sampling/sampling_logp_difference/max': 3.364473342895508, 'sampling/importance_sampling_ratio/min': 0.03458022326231003, 'sampling/importance_sampling_ratio/mean': 1.0063978433609009, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7657769918441772, 'clip_ratio/low_mean': 0.06030426826328039, 'clip_ratio/low_min': 0.06030426826328039, 'clip_ratio/high_mean': 0.05361561942845583, 'clip_ratio/high_max': 0.05361561942845583, 'clip_ratio/region_mean': 0.11391988769173622, 'reward_total_mean': 0.21185347437858582, 'reward_meter_mean': 0.7381832599639893, 'reward_meter_std': 0.2610228657722473, 'reward_count_adherence_mean': 0.8999999761581421, 'reward_count_adherence_std': 0.1511857807636261, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9304074048995972, 'reward_repeat_penalty_std': 0.13754808902740479, 'reward_near_duplicate_penalty_mean': 0.9826213121414185, 'reward_near_duplicate_penalty_std': 0.017901146784424782, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9549735188484192, 'reward_distinct_2_std': 0.1023939698934555, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.09258200973272324, 'reward_total_composite_mean': 0.21185347437858582, 'reward_total_composite_std': 0.1496955156326294, 'epoch': 0.02}

 17%|█▋        | 574/3300 [2:15:53<9:36:36, 12.69s/it]INFO 04-12 13:02:05 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 575/3300 [2:16:08<10:06:42, 13.36s/it]2026-04-12 13:02:20,662 | INFO | train_grpo_train | metrics_logged mode=train step=575

                                                       
{'loss': -0.0336, 'grad_norm': 4.962844371795654, 'learning_rate': 8.260606060606061e-06, 'num_tokens': 1153024.0, 'completions/mean_length': 103.625, 'completions/min_length': 38.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 45.28571701049805, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.5643232464790344, 'rewards/meter/std': 0.3730098009109497, '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/lexical_plausibility/mean': 0.9652791023254395, 'rewards/lexical_plausibility/std': 0.0982055589556694, 'rewards/judge_quality/mean': 0.39625000953674316, 'rewards/judge_quality/std': 0.24136146903038025, 'rewards/repeat_penalty/mean': 0.9623034596443176, 'rewards/repeat_penalty/std': 0.08640687167644501, 'rewards/near_duplicate_penalty/mean': 0.9935534596443176, 'rewards/near_duplicate_penalty/std': 0.010682421736419201, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2575927674770355, 'rewards/total_composite/std': 0.2454334944486618, 'reward': 0.2575927674770355, 'reward_std': 0.2454334795475006, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18178588151931763, 'sampling/sampling_logp_difference/max': 1.4608869552612305, 'sampling/importance_sampling_ratio/min': 0.23203037679195404, 'sampling/importance_sampling_ratio/mean': 1.0186171531677246, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0370404049754143, 'clip_ratio/low_mean': 0.04936229903250933, 'clip_ratio/low_min': 0.04936229903250933, 'clip_ratio/high_mean': 0.0777665888890624, 'clip_ratio/high_max': 0.0777665888890624, 'clip_ratio/region_mean': 0.12712888792157173, 'reward_total_mean': 0.2575927674770355, 'reward_meter_mean': 0.5643232464790344, 'reward_meter_std': 0.3730098009109497, '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_lexical_plausibility_mean': 0.9652791023254395, 'reward_lexical_plausibility_std': 0.0982055589556694, 'reward_repeat_penalty_mean': 0.9623034596443176, 'reward_repeat_penalty_std': 0.08640687167644501, 'reward_near_duplicate_penalty_mean': 0.9935534596443176, 'reward_near_duplicate_penalty_std': 0.010682421736419201, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.39625000953674316, 'reward_judge_quality_std': 0.24136146903038025, 'reward_total_composite_mean': 0.2575927674770355, 'reward_total_composite_std': 0.2454334944486618, 'epoch': 0.02}

 17%|█▋        | 575/3300 [2:16:08<10:06:42, 13.36s/it]INFO 04-12 13:02:20 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 576/3300 [2:16:24<10:42:33, 14.15s/it]2026-04-12 13:02:36,673 | INFO | train_grpo_train | metrics_logged mode=train step=576

                                                       
{'loss': -0.1999, 'grad_norm': 3.657471179962158, 'learning_rate': 8.257575757575758e-06, 'num_tokens': 1155175.0, 'completions/mean_length': 145.875, 'completions/min_length': 49.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 93.5714340209961, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 121.0, 'rewards/meter/mean': 0.940564751625061, 'rewards/meter/std': 0.06310593336820602, 'rewards/count_adherence/mean': 0.84375, 'rewards/count_adherence/std': 0.22903135418891907, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.962358832359314, 'rewards/lexical_plausibility/std': 0.10646536946296692, 'rewards/judge_quality/mean': 0.35624998807907104, 'rewards/judge_quality/std': 0.15221577882766724, 'rewards/repeat_penalty/mean': 0.9759054183959961, 'rewards/repeat_penalty/std': 0.03560880199074745, 'rewards/near_duplicate_penalty/mean': 0.9866353869438171, 'rewards/near_duplicate_penalty/std': 0.009828095324337482, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9890109896659851, 'rewards/distinct_2/std': 0.031081615015864372, 'rewards/total_composite/mean': 0.2960255444049835, 'rewards/total_composite/std': 0.1426866203546524, 'reward': 0.2960255444049835, 'reward_std': 0.1426866203546524, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15812592208385468, 'sampling/sampling_logp_difference/max': 1.9688575267791748, 'sampling/importance_sampling_ratio/min': 0.13961626589298248, 'sampling/importance_sampling_ratio/mean': 1.0087487697601318, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9720896631479263, 'clip_ratio/low_mean': 0.03464866895228624, 'clip_ratio/low_min': 0.03464866895228624, 'clip_ratio/high_mean': 0.09635142516344786, 'clip_ratio/high_max': 0.09635142516344786, 'clip_ratio/region_mean': 0.1310000941157341, 'reward_total_mean': 0.2960255444049835, 'reward_meter_mean': 0.940564751625061, 'reward_meter_std': 0.06310593336820602, 'reward_count_adherence_mean': 0.84375, 'reward_count_adherence_std': 0.22903135418891907, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.962358832359314, 'reward_lexical_plausibility_std': 0.10646536946296692, 'reward_repeat_penalty_mean': 0.9759054183959961, 'reward_repeat_penalty_std': 0.03560880199074745, 'reward_near_duplicate_penalty_mean': 0.9866353869438171, 'reward_near_duplicate_penalty_std': 0.009828095324337482, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9890109896659851, 'reward_distinct_2_std': 0.031081615015864372, 'reward_judge_quality_mean': 0.35624998807907104, 'reward_judge_quality_std': 0.15221577882766724, 'reward_total_composite_mean': 0.2960255444049835, 'reward_total_composite_std': 0.1426866203546524, 'epoch': 0.02}

 17%|█▋        | 576/3300 [2:16:24<10:42:33, 14.15s/it]INFO 04-12 13:02:36 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 577/3300 [2:16:33<9:31:27, 12.59s/it] 2026-04-12 13:02:45,616 | INFO | train_grpo_train | metrics_logged mode=train step=577

                                                      
{'loss': -0.0203, 'grad_norm': 19.71200942993164, 'learning_rate': 8.254545454545456e-06, 'num_tokens': 1156606.0, 'completions/mean_length': 21.875, 'completions/min_length': 20.0, 'completions/max_length': 25.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.875, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.7545794248580933, 'rewards/meter/std': 0.3721073269844055, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8062499761581421, 'rewards/judge_quality/std': 0.22012579441070557, 'rewards/repeat_penalty/mean': 0.7231106758117676, 'rewards/repeat_penalty/std': 0.06289298832416534, 'rewards/near_duplicate_penalty/mean': 0.9641475677490234, 'rewards/near_duplicate_penalty/std': 0.08385732024908066, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6397749185562134, 'rewards/total_composite/std': 0.3437376618385315, 'reward': 0.6397749185562134, 'reward_std': 0.3437376618385315, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11730854213237762, 'sampling/sampling_logp_difference/max': 1.0412530899047852, 'sampling/importance_sampling_ratio/min': 0.3530120551586151, 'sampling/importance_sampling_ratio/mean': 1.007327914237976, 'sampling/importance_sampling_ratio/max': 1.7369740009307861, 'entropy': 0.9567297101020813, 'clip_ratio/low_mean': 0.05338203627616167, 'clip_ratio/low_min': 0.05338203627616167, 'clip_ratio/high_mean': 0.06627435237169266, 'clip_ratio/high_max': 0.06627435237169266, 'clip_ratio/region_mean': 0.11965638864785433, 'reward_total_mean': 0.6397749185562134, 'reward_meter_mean': 0.7545794248580933, 'reward_meter_std': 0.3721073269844055, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7231106758117676, 'reward_repeat_penalty_std': 0.06289298832416534, 'reward_near_duplicate_penalty_mean': 0.9641475677490234, 'reward_near_duplicate_penalty_std': 0.08385732024908066, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8062499761581421, 'reward_judge_quality_std': 0.22012579441070557, 'reward_total_composite_mean': 0.6397749185562134, 'reward_total_composite_std': 0.3437376618385315, 'epoch': 0.02}

 17%|█▋        | 577/3300 [2:16:33<9:31:27, 12.59s/it]INFO 04-12 13:02:45 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 578/3300 [2:16:46<9:40:33, 12.80s/it]2026-04-12 13:02:58,892 | INFO | train_grpo_train | metrics_logged mode=train step=578

                                                      
{'loss': 0.0114, 'grad_norm': 3.3425655364990234, 'learning_rate': 8.251515151515153e-06, 'num_tokens': 1161338.0, 'completions/mean_length': 340.5, 'completions/min_length': 324.0, 'completions/max_length': 375.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 340.5, 'completions/min_terminated_length': 324.0, 'completions/max_terminated_length': 375.0, 'rewards/meter/mean': 0.8657701015472412, 'rewards/meter/std': 0.3493468165397644, 'rewards/count_adherence/mean': 0.8181818127632141, 'rewards/count_adherence/std': 0.048592954874038696, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.17500001192092896, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/repeat_penalty/mean': 0.008949579671025276, 'rewards/repeat_penalty/std': 0.016940230503678322, 'rewards/near_duplicate_penalty/mean': 0.5584791302680969, 'rewards/near_duplicate_penalty/std': 0.20024442672729492, 'rewards/opening_diversity/mean': 0.6278331279754639, 'rewards/opening_diversity/std': 0.3554748594760895, 'rewards/distinct_2/mean': 0.016301684081554413, 'rewards/distinct_2/std': 0.030627399682998657, 'rewards/total_composite/mean': 0.07612084597349167, 'rewards/total_composite/std': 0.06336943060159683, 'reward': 0.07612084597349167, 'reward_std': 0.06336943060159683, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04761502891778946, 'sampling/sampling_logp_difference/max': 1.5711016654968262, 'sampling/importance_sampling_ratio/min': 0.20781610906124115, 'sampling/importance_sampling_ratio/mean': 0.9998597502708435, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.31269117072224617, 'clip_ratio/low_mean': 0.014324888703413308, 'clip_ratio/low_min': 0.014324888703413308, 'clip_ratio/high_mean': 0.020633989362977445, 'clip_ratio/high_max': 0.020633989362977445, 'clip_ratio/region_mean': 0.03495887806639075, 'reward_total_mean': 0.07612084597349167, 'reward_meter_mean': 0.8657701015472412, 'reward_meter_std': 0.3493468165397644, 'reward_count_adherence_mean': 0.8181818127632141, 'reward_count_adherence_std': 0.048592954874038696, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.008949579671025276, 'reward_repeat_penalty_std': 0.016940230503678322, 'reward_near_duplicate_penalty_mean': 0.5584791302680969, 'reward_near_duplicate_penalty_std': 0.20024442672729492, 'reward_opening_diversity_mean': 0.6278331279754639, 'reward_opening_diversity_std': 0.3554748594760895, 'reward_distinct_2_mean': 0.016301684081554413, 'reward_distinct_2_std': 0.030627399682998657, 'reward_judge_quality_mean': 0.17500001192092896, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.07612084597349167, 'reward_total_composite_std': 0.06336943060159683, 'epoch': 0.02}

 18%|█▊        | 578/3300 [2:16:46<9:40:33, 12.80s/it]INFO 04-12 13:02:59 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 579/3300 [2:16:55<8:47:48, 11.64s/it]2026-04-12 13:03:07,829 | INFO | train_grpo_train | metrics_logged mode=train step=579

                                                      
{'loss': 0.0331, 'grad_norm': 15.134308815002441, 'learning_rate': 8.248484848484848e-06, 'num_tokens': 1162970.0, 'completions/mean_length': 42.0, 'completions/min_length': 38.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.0, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.9501508474349976, 'rewards/meter/std': 0.025710590183734894, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7225000262260437, 'rewards/judge_quality/std': 0.232854962348938, 'rewards/repeat_penalty/mean': 0.8606948256492615, 'rewards/repeat_penalty/std': 0.2064901739358902, 'rewards/near_duplicate_penalty/mean': 0.9639027118682861, 'rewards/near_duplicate_penalty/std': 0.05945350602269173, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9348984360694885, 'rewards/distinct_2/std': 0.10414022207260132, 'rewards/total_composite/mean': 0.6869061589241028, 'rewards/total_composite/std': 0.22407598793506622, 'reward': 0.6869061589241028, 'reward_std': 0.22407598793506622, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15086950361728668, 'sampling/sampling_logp_difference/max': 1.4813685417175293, 'sampling/importance_sampling_ratio/min': 0.2273263782262802, 'sampling/importance_sampling_ratio/mean': 1.0305646657943726, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.811005987226963, 'clip_ratio/low_mean': 0.029723837971687317, 'clip_ratio/low_min': 0.029723837971687317, 'clip_ratio/high_mean': 0.0991767467930913, 'clip_ratio/high_max': 0.0991767467930913, 'clip_ratio/region_mean': 0.12890058476477861, 'reward_total_mean': 0.6869061589241028, 'reward_meter_mean': 0.9501508474349976, 'reward_meter_std': 0.025710590183734894, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8606948256492615, 'reward_repeat_penalty_std': 0.2064901739358902, 'reward_near_duplicate_penalty_mean': 0.9639027118682861, 'reward_near_duplicate_penalty_std': 0.05945350602269173, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9348984360694885, 'reward_distinct_2_std': 0.10414022207260132, 'reward_judge_quality_mean': 0.7225000262260437, 'reward_judge_quality_std': 0.232854962348938, 'reward_total_composite_mean': 0.6869061589241028, 'reward_total_composite_std': 0.22407598793506622, 'epoch': 0.02}

 18%|█▊        | 579/3300 [2:16:55<8:47:48, 11.64s/it]INFO 04-12 13:03:08 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 580/3300 [2:17:10<9:28:53, 12.55s/it]2026-04-12 13:03:22,506 | INFO | train_grpo_train | metrics_logged mode=train step=580

                                                      
{'loss': -0.1163, 'grad_norm': 3.920473575592041, 'learning_rate': 8.245454545454546e-06, 'num_tokens': 1164801.0, 'completions/mean_length': 113.875, 'completions/min_length': 42.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 57.000003814697266, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.7830889225006104, 'rewards/meter/std': 0.3687279522418976, 'rewards/count_adherence/mean': 0.7083333730697632, 'rewards/count_adherence/std': 0.33034372329711914, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9654290676116943, 'rewards/lexical_plausibility/std': 0.09778142720460892, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.5481995940208435, 'rewards/repeat_penalty/std': 0.2665763795375824, 'rewards/near_duplicate_penalty/mean': 0.8030309677124023, 'rewards/near_duplicate_penalty/std': 0.12721829116344452, 'rewards/opening_diversity/mean': 0.84375, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.7613962888717651, 'rewards/distinct_2/std': 0.16005690395832062, 'rewards/total_composite/mean': 0.2469968944787979, 'rewards/total_composite/std': 0.1343400925397873, 'reward': 0.2469968944787979, 'reward_std': 0.1343400925397873, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1405726969242096, 'sampling/sampling_logp_difference/max': 3.062925338745117, 'sampling/importance_sampling_ratio/min': 0.04675073176622391, 'sampling/importance_sampling_ratio/mean': 1.0136158466339111, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7760870456695557, 'clip_ratio/low_mean': 0.03786821663379669, 'clip_ratio/low_min': 0.03786821663379669, 'clip_ratio/high_mean': 0.04754090262576938, 'clip_ratio/high_max': 0.04754090262576938, 'clip_ratio/region_mean': 0.08540911925956607, 'reward_total_mean': 0.2469968944787979, 'reward_meter_mean': 0.7830889225006104, 'reward_meter_std': 0.3687279522418976, 'reward_count_adherence_mean': 0.7083333730697632, 'reward_count_adherence_std': 0.33034372329711914, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9654290676116943, 'reward_lexical_plausibility_std': 0.09778142720460892, 'reward_repeat_penalty_mean': 0.5481995940208435, 'reward_repeat_penalty_std': 0.2665763795375824, 'reward_near_duplicate_penalty_mean': 0.8030309677124023, 'reward_near_duplicate_penalty_std': 0.12721829116344452, 'reward_opening_diversity_mean': 0.84375, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.7613962888717651, 'reward_distinct_2_std': 0.16005690395832062, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.2469968944787979, 'reward_total_composite_std': 0.1343400925397873, 'epoch': 0.02}

 18%|█▊        | 580/3300 [2:17:10<9:28:53, 12.55s/it]INFO 04-12 13:03:22 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 581/3300 [2:17:22<9:16:25, 12.28s/it]2026-04-12 13:03:34,148 | INFO | train_grpo_train | metrics_logged mode=train step=581

                                                      
{'loss': 0.0199, 'grad_norm': 7.384605884552002, 'learning_rate': 8.242424242424243e-06, 'num_tokens': 1167735.0, 'completions/mean_length': 161.75, 'completions/min_length': 137.0, 'completions/max_length': 184.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 161.75, 'completions/min_terminated_length': 137.0, 'completions/max_terminated_length': 184.0, 'rewards/meter/mean': 0.878890872001648, 'rewards/meter/std': 0.1957164853811264, 'rewards/count_adherence/mean': 0.9583333134651184, 'rewards/count_adherence/std': 0.07715168595314026, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.41249996423721313, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.8601372838020325, 'rewards/repeat_penalty/std': 0.09740442037582397, 'rewards/near_duplicate_penalty/mean': 0.9627166986465454, 'rewards/near_duplicate_penalty/std': 0.024184366688132286, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8917944431304932, 'rewards/distinct_2/std': 0.08167210966348648, 'rewards/total_composite/mean': 0.3456624448299408, 'rewards/total_composite/std': 0.08569075167179108, 'reward': 0.3456624448299408, 'reward_std': 0.08569075167179108, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1548485904932022, 'sampling/sampling_logp_difference/max': 2.275144577026367, 'sampling/importance_sampling_ratio/min': 0.10278204828500748, 'sampling/importance_sampling_ratio/mean': 1.0003671646118164, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.989480510354042, 'clip_ratio/low_mean': 0.04997400101274252, 'clip_ratio/low_min': 0.04997400101274252, 'clip_ratio/high_mean': 0.11123065650463104, 'clip_ratio/high_max': 0.11123065650463104, 'clip_ratio/region_mean': 0.16120465751737356, 'reward_total_mean': 0.3456624448299408, 'reward_meter_mean': 0.878890872001648, 'reward_meter_std': 0.1957164853811264, 'reward_count_adherence_mean': 0.9583333134651184, 'reward_count_adherence_std': 0.07715168595314026, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8601372838020325, 'reward_repeat_penalty_std': 0.09740442037582397, 'reward_near_duplicate_penalty_mean': 0.9627166986465454, 'reward_near_duplicate_penalty_std': 0.024184366688132286, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8917944431304932, 'reward_distinct_2_std': 0.08167210966348648, 'reward_judge_quality_mean': 0.41249996423721313, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.3456624448299408, 'reward_total_composite_std': 0.08569075167179108, 'epoch': 0.02}

 18%|█▊        | 581/3300 [2:17:22<9:16:25, 12.28s/it]INFO 04-12 13:03:34 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 582/3300 [2:17:30<8:28:29, 11.22s/it]2026-04-12 13:03:42,917 | INFO | train_grpo_train | metrics_logged mode=train step=582

                                                      
{'loss': 0.0364, 'grad_norm': 23.21431541442871, 'learning_rate': 8.23939393939394e-06, 'num_tokens': 1169235.0, 'completions/mean_length': 34.5, 'completions/min_length': 31.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.5, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.560826301574707, 'rewards/meter/std': 0.3040037453174591, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9898897409439087, 'rewards/lexical_plausibility/std': 0.028596237301826477, 'rewards/judge_quality/mean': 0.6225000023841858, 'rewards/judge_quality/std': 0.3285357356071472, 'rewards/repeat_penalty/mean': 0.9979792833328247, 'rewards/repeat_penalty/std': 0.005715449806302786, 'rewards/near_duplicate_penalty/mean': 0.9979792833328247, 'rewards/near_duplicate_penalty/std': 0.005715449806302786, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3604447543621063, 'rewards/total_composite/std': 0.2884591221809387, 'reward': 0.3604447543621063, 'reward_std': 0.2884591221809387, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13150335848331451, 'sampling/sampling_logp_difference/max': 2.0473315715789795, 'sampling/importance_sampling_ratio/min': 0.12907887995243073, 'sampling/importance_sampling_ratio/mean': 1.0005255937576294, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6319779381155968, 'clip_ratio/low_mean': 0.059326283633708954, 'clip_ratio/low_min': 0.059326283633708954, 'clip_ratio/high_mean': 0.05271501839160919, 'clip_ratio/high_max': 0.05271501839160919, 'clip_ratio/region_mean': 0.11204130202531815, 'reward_total_mean': 0.3604447543621063, 'reward_meter_mean': 0.560826301574707, 'reward_meter_std': 0.3040037453174591, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9898897409439087, 'reward_lexical_plausibility_std': 0.028596237301826477, 'reward_repeat_penalty_mean': 0.9979792833328247, 'reward_repeat_penalty_std': 0.005715449806302786, 'reward_near_duplicate_penalty_mean': 0.9979792833328247, 'reward_near_duplicate_penalty_std': 0.005715449806302786, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6225000023841858, 'reward_judge_quality_std': 0.3285357356071472, 'reward_total_composite_mean': 0.3604447543621063, 'reward_total_composite_std': 0.2884591221809387, 'epoch': 0.02}

 18%|█▊        | 582/3300 [2:17:30<8:28:29, 11.22s/it]INFO 04-12 13:03:43 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 583/3300 [2:17:44<9:03:48, 12.01s/it]2026-04-12 13:03:56,756 | INFO | train_grpo_train | metrics_logged mode=train step=583

                                                      
{'loss': -0.0551, 'grad_norm': 5.221027374267578, 'learning_rate': 8.236363636363637e-06, 'num_tokens': 1170841.0, 'completions/mean_length': 105.75, 'completions/min_length': 42.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 47.71428680419922, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.6080485582351685, 'rewards/meter/std': 0.43777281045913696, '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/lexical_plausibility/mean': 0.9689664244651794, 'rewards/lexical_plausibility/std': 0.08777620643377304, 'rewards/judge_quality/mean': 0.5425000190734863, 'rewards/judge_quality/std': 0.32805708050727844, 'rewards/repeat_penalty/mean': 0.9648683667182922, 'rewards/repeat_penalty/std': 0.08714817464351654, 'rewards/near_duplicate_penalty/mean': 0.9961183667182922, 'rewards/near_duplicate_penalty/std': 0.0077176871709525585, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3765876293182373, 'rewards/total_composite/std': 0.3583544194698334, 'reward': 0.3765876293182373, 'reward_std': 0.358354389667511, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13303391635417938, 'sampling/sampling_logp_difference/max': 1.2548370361328125, 'sampling/importance_sampling_ratio/min': 0.2851223051548004, 'sampling/importance_sampling_ratio/mean': 1.0224131345748901, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7562833800911903, 'clip_ratio/low_mean': 0.0714285746216774, 'clip_ratio/low_min': 0.0714285746216774, 'clip_ratio/high_mean': 0.07753804791718721, 'clip_ratio/high_max': 0.07753804791718721, 'clip_ratio/region_mean': 0.1489666225388646, 'reward_total_mean': 0.3765876293182373, 'reward_meter_mean': 0.6080485582351685, 'reward_meter_std': 0.43777281045913696, '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_lexical_plausibility_mean': 0.9689664244651794, 'reward_lexical_plausibility_std': 0.08777620643377304, 'reward_repeat_penalty_mean': 0.9648683667182922, 'reward_repeat_penalty_std': 0.08714817464351654, 'reward_near_duplicate_penalty_mean': 0.9961183667182922, 'reward_near_duplicate_penalty_std': 0.0077176871709525585, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5425000190734863, 'reward_judge_quality_std': 0.32805708050727844, 'reward_total_composite_mean': 0.3765876293182373, 'reward_total_composite_std': 0.3583544194698334, 'epoch': 0.02}

 18%|█▊        | 583/3300 [2:17:44<9:03:48, 12.01s/it]INFO 04-12 13:03:57 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 584/3300 [2:17:54<8:35:34, 11.39s/it]2026-04-12 13:04:06,700 | INFO | train_grpo_train | metrics_logged mode=train step=584

                                                      
{'loss': 0.0693, 'grad_norm': 10.465445518493652, 'learning_rate': 8.233333333333335e-06, 'num_tokens': 1172998.0, 'completions/mean_length': 87.625, 'completions/min_length': 78.0, 'completions/max_length': 104.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 87.625, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 104.0, 'rewards/meter/mean': 0.7624784708023071, 'rewards/meter/std': 0.3110302686691284, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.9221370220184326, 'rewards/repeat_penalty/std': 0.05366437882184982, 'rewards/near_duplicate_penalty/mean': 0.9665480852127075, 'rewards/near_duplicate_penalty/std': 0.018389509990811348, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9539777040481567, 'rewards/distinct_2/std': 0.05056172236800194, 'rewards/total_composite/mean': 0.3033396005630493, 'rewards/total_composite/std': 0.14325401186943054, 'reward': 0.3033396005630493, 'reward_std': 0.14325399696826935, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1579848676919937, 'sampling/sampling_logp_difference/max': 1.4948318004608154, 'sampling/importance_sampling_ratio/min': 0.22428633272647858, 'sampling/importance_sampling_ratio/mean': 1.0212000608444214, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1626967042684555, 'clip_ratio/low_mean': 0.0546164195984602, 'clip_ratio/low_min': 0.0546164195984602, 'clip_ratio/high_mean': 0.10806668363511562, 'clip_ratio/high_max': 0.10806668363511562, 'clip_ratio/region_mean': 0.16268310323357582, 'reward_total_mean': 0.3033396005630493, 'reward_meter_mean': 0.7624784708023071, 'reward_meter_std': 0.3110302686691284, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9221370220184326, 'reward_repeat_penalty_std': 0.05366437882184982, 'reward_near_duplicate_penalty_mean': 0.9665480852127075, 'reward_near_duplicate_penalty_std': 0.018389509990811348, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9539777040481567, 'reward_distinct_2_std': 0.05056172236800194, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.3033396005630493, 'reward_total_composite_std': 0.14325401186943054, 'epoch': 0.02}

 18%|█▊        | 584/3300 [2:17:54<8:35:34, 11.39s/it]INFO 04-12 13:04:06 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 585/3300 [2:18:04<8:13:35, 10.91s/it]2026-04-12 13:04:16,483 | INFO | train_grpo_train | metrics_logged mode=train step=585

                                                      
{'loss': 0.0247, 'grad_norm': 10.929614067077637, 'learning_rate': 8.23030303030303e-06, 'num_tokens': 1174916.0, 'completions/mean_length': 66.75, 'completions/min_length': 45.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.75, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.8810107111930847, 'rewards/meter/std': 0.1937389075756073, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4962499737739563, 'rewards/judge_quality/std': 0.17476005852222443, 'rewards/repeat_penalty/mean': 0.8048100471496582, 'rewards/repeat_penalty/std': 0.2865392863750458, 'rewards/near_duplicate_penalty/mean': 0.9030643105506897, 'rewards/near_duplicate_penalty/std': 0.12144850194454193, 'rewards/opening_diversity/mean': 0.921875, 'rewards/opening_diversity/std': 0.22097088396549225, 'rewards/distinct_2/mean': 0.9081178903579712, 'rewards/distinct_2/std': 0.1395706683397293, 'rewards/total_composite/mean': 0.4271959960460663, 'rewards/total_composite/std': 0.2150857150554657, 'reward': 0.4271959960460663, 'reward_std': 0.2150857299566269, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1364094763994217, 'sampling/sampling_logp_difference/max': 1.7006666660308838, 'sampling/importance_sampling_ratio/min': 0.18256177008152008, 'sampling/importance_sampling_ratio/mean': 1.010083556175232, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8348531275987625, 'clip_ratio/low_mean': 0.04211481101810932, 'clip_ratio/low_min': 0.04211481101810932, 'clip_ratio/high_mean': 0.09378836676478386, 'clip_ratio/high_max': 0.09378836676478386, 'clip_ratio/region_mean': 0.13590317778289318, 'reward_total_mean': 0.4271959960460663, 'reward_meter_mean': 0.8810107111930847, 'reward_meter_std': 0.1937389075756073, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8048100471496582, 'reward_repeat_penalty_std': 0.2865392863750458, 'reward_near_duplicate_penalty_mean': 0.9030643105506897, 'reward_near_duplicate_penalty_std': 0.12144850194454193, 'reward_opening_diversity_mean': 0.921875, 'reward_opening_diversity_std': 0.22097088396549225, 'reward_distinct_2_mean': 0.9081178903579712, 'reward_distinct_2_std': 0.1395706683397293, 'reward_judge_quality_mean': 0.4962499737739563, 'reward_judge_quality_std': 0.17476005852222443, 'reward_total_composite_mean': 0.4271959960460663, 'reward_total_composite_std': 0.2150857150554657, 'epoch': 0.02}

 18%|█▊        | 585/3300 [2:18:04<8:13:35, 10.91s/it]INFO 04-12 13:04:16 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 586/3300 [2:18:18<9:00:54, 11.96s/it]2026-04-12 13:04:30,892 | INFO | train_grpo_train | metrics_logged mode=train step=586

                                                      
{'loss': -0.0461, 'grad_norm': 4.653808116912842, 'learning_rate': 8.227272727272728e-06, 'num_tokens': 1176635.0, 'completions/mean_length': 110.875, 'completions/min_length': 48.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 53.57143020629883, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.8222522735595703, 'rewards/meter/std': 0.27135413885116577, '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/lexical_plausibility/mean': 0.9601929187774658, 'rewards/lexical_plausibility/std': 0.07594508677721024, 'rewards/judge_quality/mean': 0.5325000286102295, 'rewards/judge_quality/std': 0.34656479954719543, 'rewards/repeat_penalty/mean': 0.9464002847671509, 'rewards/repeat_penalty/std': 0.09431640803813934, 'rewards/near_duplicate_penalty/mean': 0.9856500625610352, 'rewards/near_duplicate_penalty/std': 0.02964356541633606, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9912587404251099, 'rewards/distinct_2/std': 0.024724015966057777, 'rewards/total_composite/mean': 0.46150755882263184, 'rewards/total_composite/std': 0.36525657773017883, 'reward': 0.46150755882263184, 'reward_std': 0.36525657773017883, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17011520266532898, 'sampling/sampling_logp_difference/max': 2.3689229488372803, 'sampling/importance_sampling_ratio/min': 0.09358146041631699, 'sampling/importance_sampling_ratio/mean': 1.0176448822021484, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.152994766831398, 'clip_ratio/low_mean': 0.08484716154634953, 'clip_ratio/low_min': 0.08484716154634953, 'clip_ratio/high_mean': 0.06638168357312679, 'clip_ratio/high_max': 0.06638168357312679, 'clip_ratio/region_mean': 0.15122884511947632, 'reward_total_mean': 0.46150755882263184, 'reward_meter_mean': 0.8222522735595703, 'reward_meter_std': 0.27135413885116577, '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_lexical_plausibility_mean': 0.9601929187774658, 'reward_lexical_plausibility_std': 0.07594508677721024, 'reward_repeat_penalty_mean': 0.9464002847671509, 'reward_repeat_penalty_std': 0.09431640803813934, 'reward_near_duplicate_penalty_mean': 0.9856500625610352, 'reward_near_duplicate_penalty_std': 0.02964356541633606, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9912587404251099, 'reward_distinct_2_std': 0.024724015966057777, 'reward_judge_quality_mean': 0.5325000286102295, 'reward_judge_quality_std': 0.34656479954719543, 'reward_total_composite_mean': 0.46150755882263184, 'reward_total_composite_std': 0.36525657773017883, 'epoch': 0.02}

 18%|█▊        | 586/3300 [2:18:18<9:00:54, 11.96s/it]INFO 04-12 13:04:31 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 587/3300 [2:18:32<9:29:20, 12.59s/it]2026-04-12 13:04:44,960 | INFO | train_grpo_train | metrics_logged mode=train step=587

                                                      
{'loss': -0.0732, 'grad_norm': 5.389424800872803, 'learning_rate': 8.224242424242425e-06, 'num_tokens': 1178294.0, 'completions/mean_length': 108.375, 'completions/min_length': 44.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 50.71428680419922, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.9181538820266724, 'rewards/meter/std': 0.1899999976158142, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.95954430103302, 'rewards/lexical_plausibility/std': 0.0839536190032959, 'rewards/judge_quality/mean': 0.42500001192092896, 'rewards/judge_quality/std': 0.19086270034313202, 'rewards/repeat_penalty/mean': 0.9667361378669739, 'rewards/repeat_penalty/std': 0.04056604579091072, 'rewards/near_duplicate_penalty/mean': 0.9820794463157654, 'rewards/near_duplicate_penalty/std': 0.017676454037427902, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9841362237930298, 'rewards/distinct_2/std': 0.029577190056443214, 'rewards/total_composite/mean': 0.38899046182632446, 'rewards/total_composite/std': 0.20348544418811798, 'reward': 0.38899046182632446, 'reward_std': 0.2034854143857956, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18173500895500183, 'sampling/sampling_logp_difference/max': 1.330775260925293, 'sampling/importance_sampling_ratio/min': 0.2642723023891449, 'sampling/importance_sampling_ratio/mean': 1.035598874092102, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.385396733880043, 'clip_ratio/low_mean': 0.037711864337325096, 'clip_ratio/low_min': 0.037711864337325096, 'clip_ratio/high_mean': 0.12368344143033028, 'clip_ratio/high_max': 0.12368344143033028, 'clip_ratio/region_mean': 0.16139530576765537, 'reward_total_mean': 0.38899046182632446, 'reward_meter_mean': 0.9181538820266724, 'reward_meter_std': 0.1899999976158142, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.95954430103302, 'reward_lexical_plausibility_std': 0.0839536190032959, 'reward_repeat_penalty_mean': 0.9667361378669739, 'reward_repeat_penalty_std': 0.04056604579091072, 'reward_near_duplicate_penalty_mean': 0.9820794463157654, 'reward_near_duplicate_penalty_std': 0.017676454037427902, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9841362237930298, 'reward_distinct_2_std': 0.029577190056443214, 'reward_judge_quality_mean': 0.42500001192092896, 'reward_judge_quality_std': 0.19086270034313202, 'reward_total_composite_mean': 0.38899046182632446, 'reward_total_composite_std': 0.20348544418811798, 'epoch': 0.02}

 18%|█▊        | 587/3300 [2:18:32<9:29:20, 12.59s/it]INFO 04-12 13:04:45 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 588/3300 [2:18:44<9:12:58, 12.23s/it]2026-04-12 13:04:56,359 | INFO | train_grpo_train | metrics_logged mode=train step=588

                                                      
{'loss': 0.0638, 'grad_norm': 3.340954065322876, 'learning_rate': 8.221212121212122e-06, 'num_tokens': 1181666.0, 'completions/mean_length': 230.5, 'completions/min_length': 203.0, 'completions/max_length': 263.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 230.5, 'completions/min_terminated_length': 203.0, 'completions/max_terminated_length': 263.0, 'rewards/meter/mean': 0.9035285711288452, 'rewards/meter/std': 0.18135115504264832, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.0534522607922554, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.0020673582330346107, 'rewards/repeat_penalty/std': 0.005847372114658356, 'rewards/near_duplicate_penalty/mean': 0.47508320212364197, 'rewards/near_duplicate_penalty/std': 0.13144657015800476, 'rewards/opening_diversity/mean': 0.47642046213150024, 'rewards/opening_diversity/std': 0.33804193139076233, 'rewards/distinct_2/mean': 0.0158083438873291, 'rewards/distinct_2/std': 0.044712748378515244, 'rewards/total_composite/mean': 0.20263038575649261, 'rewards/total_composite/std': 0.11826668679714203, 'reward': 0.20263038575649261, 'reward_std': 0.11826668679714203, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0420808270573616, 'sampling/sampling_logp_difference/max': 1.8253421783447266, 'sampling/importance_sampling_ratio/min': 0.16116249561309814, 'sampling/importance_sampling_ratio/mean': 1.0034795999526978, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.27530100382864475, 'clip_ratio/low_mean': 0.01830257789697498, 'clip_ratio/low_min': 0.01830257789697498, 'clip_ratio/high_mean': 0.018705840688198805, 'clip_ratio/high_max': 0.018705840688198805, 'clip_ratio/region_mean': 0.037008418585173786, 'reward_total_mean': 0.20263038575649261, 'reward_meter_mean': 0.9035285711288452, 'reward_meter_std': 0.18135115504264832, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.0534522607922554, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.0020673582330346107, 'reward_repeat_penalty_std': 0.005847372114658356, 'reward_near_duplicate_penalty_mean': 0.47508320212364197, 'reward_near_duplicate_penalty_std': 0.13144657015800476, 'reward_opening_diversity_mean': 0.47642046213150024, 'reward_opening_diversity_std': 0.33804193139076233, 'reward_distinct_2_mean': 0.0158083438873291, 'reward_distinct_2_std': 0.044712748378515244, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.20263038575649261, 'reward_total_composite_std': 0.11826668679714203, 'epoch': 0.02}

 18%|█▊        | 588/3300 [2:18:44<9:12:58, 12.23s/it]INFO 04-12 13:04:56 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 589/3300 [2:18:58<9:41:11, 12.86s/it]2026-04-12 13:05:10,701 | INFO | train_grpo_train | metrics_logged mode=train step=589

                                                      
{'loss': -0.0561, 'grad_norm': 6.356830596923828, 'learning_rate': 8.21818181818182e-06, 'num_tokens': 1183411.0, 'completions/mean_length': 113.125, 'completions/min_length': 47.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 56.142860412597656, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.48142874240875244, 'rewards/meter/std': 0.30561205744743347, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9659296870231628, 'rewards/lexical_plausibility/std': 0.09636537730693817, 'rewards/judge_quality/mean': 0.3375000059604645, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.9884310960769653, 'rewards/repeat_penalty/std': 0.018634436652064323, 'rewards/near_duplicate_penalty/mean': 0.9884310960769653, 'rewards/near_duplicate_penalty/std': 0.018634436652064323, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1727956235408783, 'rewards/total_composite/std': 0.1319209635257721, 'reward': 0.1727956235408783, 'reward_std': 0.1319209635257721, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1446923166513443, 'sampling/sampling_logp_difference/max': 1.9724135398864746, 'sampling/importance_sampling_ratio/min': 0.1391206830739975, 'sampling/importance_sampling_ratio/mean': 0.9802079796791077, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6234605275094509, 'clip_ratio/low_mean': 0.025961538776755333, 'clip_ratio/low_min': 0.025961538776755333, 'clip_ratio/high_mean': 0.08062838343903422, 'clip_ratio/high_max': 0.08062838343903422, 'clip_ratio/region_mean': 0.10658992221578956, 'reward_total_mean': 0.1727956235408783, 'reward_meter_mean': 0.48142874240875244, 'reward_meter_std': 0.30561205744743347, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9659296870231628, 'reward_lexical_plausibility_std': 0.09636537730693817, 'reward_repeat_penalty_mean': 0.9884310960769653, 'reward_repeat_penalty_std': 0.018634436652064323, 'reward_near_duplicate_penalty_mean': 0.9884310960769653, 'reward_near_duplicate_penalty_std': 0.018634436652064323, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3375000059604645, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.1727956235408783, 'reward_total_composite_std': 0.1319209635257721, 'epoch': 0.02}

 18%|█▊        | 589/3300 [2:18:58<9:41:11, 12.86s/it]INFO 04-12 13:05:10 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 590/3300 [2:19:10<9:24:57, 12.51s/it]2026-04-12 13:05:22,370 | INFO | train_grpo_train | metrics_logged mode=train step=590

                                                      
{'loss': 0.0032, 'grad_norm': 5.88330602645874, 'learning_rate': 8.215151515151517e-06, 'num_tokens': 1186330.0, 'completions/mean_length': 174.875, 'completions/min_length': 145.0, 'completions/max_length': 200.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 174.875, 'completions/min_terminated_length': 145.0, 'completions/max_terminated_length': 200.0, 'rewards/meter/mean': 0.9274966716766357, 'rewards/meter/std': 0.07063567638397217, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0578637570142746, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.07253516465425491, 'rewards/repeat_penalty/std': 0.08631241321563721, 'rewards/near_duplicate_penalty/mean': 0.7592816352844238, 'rewards/near_duplicate_penalty/std': 0.09269137680530548, 'rewards/opening_diversity/mean': 0.7142857313156128, 'rewards/opening_diversity/std': 0.23006081581115723, 'rewards/distinct_2/mean': 0.15668581426143646, 'rewards/distinct_2/std': 0.19408312439918518, 'rewards/total_composite/mean': 0.1346191167831421, 'rewards/total_composite/std': 0.011395241133868694, 'reward': 0.1346191167831421, 'reward_std': 0.01139524020254612, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07822994142770767, 'sampling/sampling_logp_difference/max': 2.2538514137268066, 'sampling/importance_sampling_ratio/min': 0.1437089443206787, 'sampling/importance_sampling_ratio/mean': 1.0050102472305298, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4841975122690201, 'clip_ratio/low_mean': 0.03451812267303467, 'clip_ratio/low_min': 0.03451812267303467, 'clip_ratio/high_mean': 0.03097512573003769, 'clip_ratio/high_max': 0.03097512573003769, 'clip_ratio/region_mean': 0.06549324840307236, 'reward_total_mean': 0.1346191167831421, 'reward_meter_mean': 0.9274966716766357, 'reward_meter_std': 0.07063567638397217, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0578637570142746, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.07253516465425491, 'reward_repeat_penalty_std': 0.08631241321563721, 'reward_near_duplicate_penalty_mean': 0.7592816352844238, 'reward_near_duplicate_penalty_std': 0.09269137680530548, 'reward_opening_diversity_mean': 0.7142857313156128, 'reward_opening_diversity_std': 0.23006081581115723, 'reward_distinct_2_mean': 0.15668581426143646, 'reward_distinct_2_std': 0.19408312439918518, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.1346191167831421, 'reward_total_composite_std': 0.011395241133868694, 'epoch': 0.02}

 18%|█▊        | 590/3300 [2:19:10<9:24:57, 12.51s/it]INFO 04-12 13:05:22 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 591/3300 [2:19:19<8:44:21, 11.61s/it]2026-04-12 13:05:31,896 | INFO | train_grpo_train | metrics_logged mode=train step=591

                                                      
{'loss': 0.0358, 'grad_norm': 11.83116626739502, 'learning_rate': 8.212121212121212e-06, 'num_tokens': 1188316.0, 'completions/mean_length': 73.25, 'completions/min_length': 68.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.25, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.6335621476173401, 'rewards/meter/std': 0.3610627055168152, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4712499976158142, 'rewards/judge_quality/std': 0.18795421719551086, 'rewards/repeat_penalty/mean': 0.9570281505584717, 'rewards/repeat_penalty/std': 0.052605655044317245, 'rewards/near_duplicate_penalty/mean': 0.9728123545646667, 'rewards/near_duplicate_penalty/std': 0.01498380582779646, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9835164546966553, 'rewards/distinct_2/std': 0.046622421592473984, 'rewards/total_composite/mean': 0.2758263945579529, 'rewards/total_composite/std': 0.13896574079990387, 'reward': 0.2758263945579529, 'reward_std': 0.13896572589874268, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17875981330871582, 'sampling/sampling_logp_difference/max': 1.8909282684326172, 'sampling/importance_sampling_ratio/min': 0.15093164145946503, 'sampling/importance_sampling_ratio/mean': 1.0222339630126953, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4100866466760635, 'clip_ratio/low_mean': 0.07188650220632553, 'clip_ratio/low_min': 0.07188650220632553, 'clip_ratio/high_mean': 0.07916142418980598, 'clip_ratio/high_max': 0.07916142418980598, 'clip_ratio/region_mean': 0.15104792639613152, 'reward_total_mean': 0.2758263945579529, 'reward_meter_mean': 0.6335621476173401, 'reward_meter_std': 0.3610627055168152, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9570281505584717, 'reward_repeat_penalty_std': 0.052605655044317245, 'reward_near_duplicate_penalty_mean': 0.9728123545646667, 'reward_near_duplicate_penalty_std': 0.01498380582779646, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9835164546966553, 'reward_distinct_2_std': 0.046622421592473984, 'reward_judge_quality_mean': 0.4712499976158142, 'reward_judge_quality_std': 0.18795421719551086, 'reward_total_composite_mean': 0.2758263945579529, 'reward_total_composite_std': 0.13896574079990387, 'epoch': 0.02}

 18%|█▊        | 591/3300 [2:19:19<8:44:21, 11.61s/it]INFO 04-12 13:05:32 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 592/3300 [2:19:30<8:24:46, 11.18s/it]2026-04-12 13:05:42,107 | INFO | train_grpo_train | metrics_logged mode=train step=592

                                                      
{'loss': -0.002, 'grad_norm': 9.355712890625, 'learning_rate': 8.20909090909091e-06, 'num_tokens': 1190774.0, 'completions/mean_length': 108.25, 'completions/min_length': 91.0, 'completions/max_length': 115.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 108.25, 'completions/min_terminated_length': 91.0, 'completions/max_terminated_length': 115.0, 'rewards/meter/mean': 0.6430141925811768, 'rewards/meter/std': 0.26479583978652954, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/repeat_penalty/mean': 0.46775728464126587, 'rewards/repeat_penalty/std': 0.2647733688354492, 'rewards/near_duplicate_penalty/mean': 0.8557314872741699, 'rewards/near_duplicate_penalty/std': 0.1383117288351059, 'rewards/opening_diversity/mean': 0.729687511920929, 'rewards/opening_diversity/std': 0.25477033853530884, 'rewards/distinct_2/mean': 0.6347394585609436, 'rewards/distinct_2/std': 0.24497342109680176, 'rewards/total_composite/mean': 0.19323819875717163, 'rewards/total_composite/std': 0.1093558818101883, 'reward': 0.19323819875717163, 'reward_std': 0.10935588926076889, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12800531089305878, 'sampling/sampling_logp_difference/max': 1.877915382385254, 'sampling/importance_sampling_ratio/min': 0.1529085338115692, 'sampling/importance_sampling_ratio/mean': 1.002397060394287, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8475909605622292, 'clip_ratio/low_mean': 0.055872602853924036, 'clip_ratio/low_min': 0.055872602853924036, 'clip_ratio/high_mean': 0.0824524313211441, 'clip_ratio/high_max': 0.0824524313211441, 'clip_ratio/region_mean': 0.13832503417506814, 'reward_total_mean': 0.19323819875717163, 'reward_meter_mean': 0.6430141925811768, 'reward_meter_std': 0.26479583978652954, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.46775728464126587, 'reward_repeat_penalty_std': 0.2647733688354492, 'reward_near_duplicate_penalty_mean': 0.8557314872741699, 'reward_near_duplicate_penalty_std': 0.1383117288351059, 'reward_opening_diversity_mean': 0.729687511920929, 'reward_opening_diversity_std': 0.25477033853530884, 'reward_distinct_2_mean': 0.6347394585609436, 'reward_distinct_2_std': 0.24497342109680176, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.19323819875717163, 'reward_total_composite_std': 0.1093558818101883, 'epoch': 0.02}

 18%|█▊        | 592/3300 [2:19:30<8:24:46, 11.18s/it]INFO 04-12 13:05:42 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 593/3300 [2:19:41<8:31:44, 11.34s/it]2026-04-12 13:05:53,798 | INFO | train_grpo_train | metrics_logged mode=train step=593

                                                      
{'loss': 0.0391, 'grad_norm': 5.09604024887085, 'learning_rate': 8.206060606060607e-06, 'num_tokens': 1193512.0, 'completions/mean_length': 158.25, 'completions/min_length': 140.0, 'completions/max_length': 173.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 158.25, 'completions/min_terminated_length': 140.0, 'completions/max_terminated_length': 173.0, 'rewards/meter/mean': 0.9881192445755005, 'rewards/meter/std': 0.004923494532704353, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.10690449178218842, 'rewards/repeat_penalty/mean': 0.0976167768239975, 'rewards/repeat_penalty/std': 0.12892010807991028, 'rewards/near_duplicate_penalty/mean': 0.6389048099517822, 'rewards/near_duplicate_penalty/std': 0.16052363812923431, 'rewards/opening_diversity/mean': 0.65625, 'rewards/opening_diversity/std': 0.22903135418891907, 'rewards/distinct_2/mean': 0.1999150812625885, 'rewards/distinct_2/std': 0.1849735677242279, 'rewards/total_composite/mean': 0.24686813354492188, 'rewards/total_composite/std': 0.10521165281534195, 'reward': 0.24686813354492188, 'reward_std': 0.10521166771650314, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06000611186027527, 'sampling/sampling_logp_difference/max': 1.356687068939209, 'sampling/importance_sampling_ratio/min': 0.25751248002052307, 'sampling/importance_sampling_ratio/mean': 1.0063353776931763, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4491439461708069, 'clip_ratio/low_mean': 0.018415135331451893, 'clip_ratio/low_min': 0.018415135331451893, 'clip_ratio/high_mean': 0.03881014045327902, 'clip_ratio/high_max': 0.03881014045327902, 'clip_ratio/region_mean': 0.05722527578473091, 'reward_total_mean': 0.24686813354492188, 'reward_meter_mean': 0.9881192445755005, 'reward_meter_std': 0.004923494532704353, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.0976167768239975, 'reward_repeat_penalty_std': 0.12892010807991028, 'reward_near_duplicate_penalty_mean': 0.6389048099517822, 'reward_near_duplicate_penalty_std': 0.16052363812923431, 'reward_opening_diversity_mean': 0.65625, 'reward_opening_diversity_std': 0.22903135418891907, 'reward_distinct_2_mean': 0.1999150812625885, 'reward_distinct_2_std': 0.1849735677242279, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.10690449178218842, 'reward_total_composite_mean': 0.24686813354492188, 'reward_total_composite_std': 0.10521165281534195, 'epoch': 0.02}

 18%|█▊        | 593/3300 [2:19:41<8:31:44, 11.34s/it]INFO 04-12 13:05:54 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 594/3300 [2:19:55<9:07:32, 12.14s/it]2026-04-12 13:06:07,795 | INFO | train_grpo_train | metrics_logged mode=train step=594

                                                      
{'loss': -0.0704, 'grad_norm': 2.8470280170440674, 'learning_rate': 8.203030303030304e-06, 'num_tokens': 1195056.0, 'completions/mean_length': 222.0, 'completions/min_length': 41.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 48.0, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.4698549509048462, 'rewards/meter/std': 0.4591921865940094, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.8958449959754944, 'rewards/lexical_plausibility/std': 0.12080350518226624, 'rewards/judge_quality/mean': 0.28125, 'rewards/judge_quality/std': 0.2086307406425476, 'rewards/repeat_penalty/mean': 0.9939273595809937, 'rewards/repeat_penalty/std': 0.013081302866339684, 'rewards/near_duplicate_penalty/mean': 0.9939273595809937, 'rewards/near_duplicate_penalty/std': 0.013081302866339684, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.13311700522899628, 'rewards/total_composite/std': 0.18157075345516205, 'reward': 0.13311700522899628, 'reward_std': 0.18157075345516205, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18174710869789124, 'sampling/sampling_logp_difference/max': 1.377575159072876, 'sampling/importance_sampling_ratio/min': 0.2521893382072449, 'sampling/importance_sampling_ratio/mean': 1.034491777420044, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9815448224544525, 'clip_ratio/low_mean': 0.030424287542700768, 'clip_ratio/low_min': 0.030424287542700768, 'clip_ratio/high_mean': 0.06478696875274181, 'clip_ratio/high_max': 0.06478696875274181, 'clip_ratio/region_mean': 0.09521125629544258, 'reward_total_mean': 0.13311700522899628, 'reward_meter_mean': 0.4698549509048462, 'reward_meter_std': 0.4591921865940094, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.8958449959754944, 'reward_lexical_plausibility_std': 0.12080350518226624, 'reward_repeat_penalty_mean': 0.9939273595809937, 'reward_repeat_penalty_std': 0.013081302866339684, 'reward_near_duplicate_penalty_mean': 0.9939273595809937, 'reward_near_duplicate_penalty_std': 0.013081302866339684, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.28125, 'reward_judge_quality_std': 0.2086307406425476, 'reward_total_composite_mean': 0.13311700522899628, 'reward_total_composite_std': 0.18157075345516205, 'epoch': 0.02}

 18%|█▊        | 594/3300 [2:19:55<9:07:32, 12.14s/it]INFO 04-12 13:06:08 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 595/3300 [2:20:05<8:40:58, 11.56s/it]2026-04-12 13:06:17,989 | INFO | train_grpo_train | metrics_logged mode=train step=595

                                                      
{'loss': 0.013, 'grad_norm': 8.841227531433105, 'learning_rate': 8.2e-06, 'num_tokens': 1197953.0, 'completions/mean_length': 142.125, 'completions/min_length': 127.0, 'completions/max_length': 151.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 142.125, 'completions/min_terminated_length': 127.0, 'completions/max_terminated_length': 151.0, 'rewards/meter/mean': 0.9749481678009033, 'rewards/meter/std': 0.014264386147260666, 'rewards/count_adherence/mean': 0.9791666269302368, 'rewards/count_adherence/std': 0.0589255727827549, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/repeat_penalty/mean': 0.7836911678314209, 'rewards/repeat_penalty/std': 0.09300744533538818, 'rewards/near_duplicate_penalty/mean': 0.9508271217346191, 'rewards/near_duplicate_penalty/std': 0.0181761272251606, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0578637570142746, 'rewards/distinct_2/mean': 0.8482184410095215, 'rewards/distinct_2/std': 0.050839297473430634, 'rewards/total_composite/mean': 0.35877329111099243, 'rewards/total_composite/std': 0.0758342295885086, 'reward': 0.35877329111099243, 'reward_std': 0.0758342295885086, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14057113230228424, 'sampling/sampling_logp_difference/max': 3.4908461570739746, 'sampling/importance_sampling_ratio/min': 0.03047507256269455, 'sampling/importance_sampling_ratio/mean': 1.0044574737548828, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8607720509171486, 'clip_ratio/low_mean': 0.07302203122526407, 'clip_ratio/low_min': 0.07302203122526407, 'clip_ratio/high_mean': 0.05689534358680248, 'clip_ratio/high_max': 0.05689534358680248, 'clip_ratio/region_mean': 0.12991737481206656, 'reward_total_mean': 0.35877329111099243, 'reward_meter_mean': 0.9749481678009033, 'reward_meter_std': 0.014264386147260666, 'reward_count_adherence_mean': 0.9791666269302368, 'reward_count_adherence_std': 0.0589255727827549, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7836911678314209, 'reward_repeat_penalty_std': 0.09300744533538818, 'reward_near_duplicate_penalty_mean': 0.9508271217346191, 'reward_near_duplicate_penalty_std': 0.0181761272251606, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0578637570142746, 'reward_distinct_2_mean': 0.8482184410095215, 'reward_distinct_2_std': 0.050839297473430634, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.35877329111099243, 'reward_total_composite_std': 0.0758342295885086, 'epoch': 0.02}

 18%|█▊        | 595/3300 [2:20:05<8:40:58, 11.56s/it]INFO 04-12 13:06:18 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 596/3300 [2:20:15<8:09:32, 10.86s/it]2026-04-12 13:06:27,231 | INFO | train_grpo_train | metrics_logged mode=train step=596

                                                      
{'loss': -0.0029, 'grad_norm': 11.209979057312012, 'learning_rate': 8.196969696969698e-06, 'num_tokens': 1199548.0, 'completions/mean_length': 46.375, 'completions/min_length': 40.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.375, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.8741564750671387, 'rewards/meter/std': 0.21297281980514526, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7300000190734863, 'rewards/judge_quality/std': 0.2331155687570572, 'rewards/repeat_penalty/mean': 0.7269378900527954, 'rewards/repeat_penalty/std': 0.2766073942184448, 'rewards/near_duplicate_penalty/mean': 0.9295974969863892, 'rewards/near_duplicate_penalty/std': 0.045377690345048904, 'rewards/opening_diversity/mean': 0.828125, 'rewards/opening_diversity/std': 0.22097088396549225, 'rewards/distinct_2/mean': 0.9116463661193848, 'rewards/distinct_2/std': 0.09794268012046814, 'rewards/total_composite/mean': 0.6433532238006592, 'rewards/total_composite/std': 0.276895672082901, 'reward': 0.6433532238006592, 'reward_std': 0.2768957018852234, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10909253358840942, 'sampling/sampling_logp_difference/max': 2.1995065212249756, 'sampling/importance_sampling_ratio/min': 0.11085785180330276, 'sampling/importance_sampling_ratio/mean': 0.9987157583236694, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8462506011128426, 'clip_ratio/low_mean': 0.03605072549544275, 'clip_ratio/low_min': 0.03605072549544275, 'clip_ratio/high_mean': 0.07286687940359116, 'clip_ratio/high_max': 0.07286687940359116, 'clip_ratio/region_mean': 0.1089176048990339, 'reward_total_mean': 0.6433532238006592, 'reward_meter_mean': 0.8741564750671387, 'reward_meter_std': 0.21297281980514526, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7269378900527954, 'reward_repeat_penalty_std': 0.2766073942184448, 'reward_near_duplicate_penalty_mean': 0.9295974969863892, 'reward_near_duplicate_penalty_std': 0.045377690345048904, 'reward_opening_diversity_mean': 0.828125, 'reward_opening_diversity_std': 0.22097088396549225, 'reward_distinct_2_mean': 0.9116463661193848, 'reward_distinct_2_std': 0.09794268012046814, 'reward_judge_quality_mean': 0.7300000190734863, 'reward_judge_quality_std': 0.2331155687570572, 'reward_total_composite_mean': 0.6433532238006592, 'reward_total_composite_std': 0.276895672082901, 'epoch': 0.02}

 18%|█▊        | 596/3300 [2:20:15<8:09:32, 10.86s/it]INFO 04-12 13:06:27 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 597/3300 [2:20:28<8:42:56, 11.61s/it]2026-04-12 13:06:40,588 | INFO | train_grpo_train | metrics_logged mode=train step=597

                                                      
{'loss': -0.0327, 'grad_norm': 2.5110538005828857, 'learning_rate': 8.193939393939394e-06, 'num_tokens': 1200851.0, 'completions/mean_length': 144.875, 'completions/min_length': 20.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 22.5, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.6916502714157104, 'rewards/meter/std': 0.43547436594963074, '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/lexical_plausibility/mean': 0.8985511660575867, 'rewards/lexical_plausibility/std': 0.19062189757823944, 'rewards/judge_quality/mean': 0.2587500214576721, 'rewards/judge_quality/std': 0.2947850525379181, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.22420215606689453, 'rewards/total_composite/std': 0.2921702265739441, 'reward': 0.22420215606689453, 'reward_std': 0.2921702265739441, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1687224805355072, 'sampling/sampling_logp_difference/max': 1.0874176025390625, 'sampling/importance_sampling_ratio/min': 0.33708587288856506, 'sampling/importance_sampling_ratio/mean': 1.054047703742981, 'sampling/importance_sampling_ratio/max': 1.8535075187683105, 'entropy': 1.0874253511428833, 'clip_ratio/low_mean': 0.09645562805235386, 'clip_ratio/low_min': 0.09645562805235386, 'clip_ratio/high_mean': 0.048958332277834415, 'clip_ratio/high_max': 0.048958332277834415, 'clip_ratio/region_mean': 0.14541396033018827, 'reward_total_mean': 0.22420215606689453, 'reward_meter_mean': 0.6916502714157104, 'reward_meter_std': 0.43547436594963074, '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_lexical_plausibility_mean': 0.8985511660575867, 'reward_lexical_plausibility_std': 0.19062189757823944, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.2587500214576721, 'reward_judge_quality_std': 0.2947850525379181, 'reward_total_composite_mean': 0.22420215606689453, 'reward_total_composite_std': 0.2921702265739441, 'epoch': 0.02}

 18%|█▊        | 597/3300 [2:20:28<8:42:56, 11.61s/it]INFO 04-12 13:06:40 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 598/3300 [2:20:42<9:16:29, 12.36s/it]2026-04-12 13:06:54,689 | INFO | train_grpo_train | metrics_logged mode=train step=598

                                                      
{'loss': -0.232, 'grad_norm': 2.3085203170776367, 'learning_rate': 8.190909090909091e-06, 'num_tokens': 1203279.0, 'completions/mean_length': 228.5, 'completions/min_length': 126.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 134.0, 'completions/min_terminated_length': 126.0, 'completions/max_terminated_length': 149.0, 'rewards/meter/mean': 0.86614990234375, 'rewards/meter/std': 0.3320692181587219, 'rewards/count_adherence/mean': 0.8500000238418579, 'rewards/count_adherence/std': 0.2976095378398895, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9482486248016357, 'rewards/lexical_plausibility/std': 0.09619145840406418, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.1752549111843109, 'rewards/repeat_penalty/mean': 0.8865271806716919, 'rewards/repeat_penalty/std': 0.14012016355991364, 'rewards/near_duplicate_penalty/mean': 0.969161331653595, 'rewards/near_duplicate_penalty/std': 0.03032568469643593, 'rewards/opening_diversity/mean': 0.987500011920929, 'rewards/opening_diversity/std': 0.0353553481400013, 'rewards/distinct_2/mean': 0.9205664992332458, 'rewards/distinct_2/std': 0.09932316094636917, 'rewards/total_composite/mean': 0.3072720766067505, 'rewards/total_composite/std': 0.19429200887680054, 'reward': 0.3072720766067505, 'reward_std': 0.19429200887680054, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14834855496883392, 'sampling/sampling_logp_difference/max': 1.7641253471374512, 'sampling/importance_sampling_ratio/min': 0.1713365763425827, 'sampling/importance_sampling_ratio/mean': 1.0215569734573364, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9741218537092209, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.10811312682926655, 'clip_ratio/high_max': 0.10811312682926655, 'clip_ratio/region_mean': 0.10811312682926655, 'reward_total_mean': 0.3072720766067505, 'reward_meter_mean': 0.86614990234375, 'reward_meter_std': 0.3320692181587219, 'reward_count_adherence_mean': 0.8500000238418579, 'reward_count_adherence_std': 0.2976095378398895, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9482486248016357, 'reward_lexical_plausibility_std': 0.09619145840406418, 'reward_repeat_penalty_mean': 0.8865271806716919, 'reward_repeat_penalty_std': 0.14012016355991364, 'reward_near_duplicate_penalty_mean': 0.969161331653595, 'reward_near_duplicate_penalty_std': 0.03032568469643593, 'reward_opening_diversity_mean': 0.987500011920929, 'reward_opening_diversity_std': 0.0353553481400013, 'reward_distinct_2_mean': 0.9205664992332458, 'reward_distinct_2_std': 0.09932316094636917, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.1752549111843109, 'reward_total_composite_mean': 0.3072720766067505, 'reward_total_composite_std': 0.19429200887680054, 'epoch': 0.02}

 18%|█▊        | 598/3300 [2:20:42<9:16:29, 12.36s/it]INFO 04-12 13:06:54 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 599/3300 [2:20:57<9:50:35, 13.12s/it]2026-04-12 13:07:09,589 | INFO | train_grpo_train | metrics_logged mode=train step=599

                                                      
{'loss': -0.3541, 'grad_norm': 0.9565184712409973, 'learning_rate': 8.187878787878788e-06, 'num_tokens': 1207013.0, 'completions/mean_length': 391.75, 'completions/min_length': 316.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 351.66668701171875, 'completions/min_terminated_length': 316.0, 'completions/max_terminated_length': 406.0, 'rewards/meter/mean': 0.7749155759811401, 'rewards/meter/std': 0.36531969904899597, 'rewards/count_adherence/mean': 0.6607142686843872, 'rewards/count_adherence/std': 0.36770108342170715, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9490639567375183, 'rewards/lexical_plausibility/std': 0.0980229303240776, 'rewards/judge_quality/mean': 0.125, 'rewards/judge_quality/std': 0.046291008591651917, 'rewards/repeat_penalty/mean': 0.21573305130004883, 'rewards/repeat_penalty/std': 0.4061233401298523, 'rewards/near_duplicate_penalty/mean': 0.4996299147605896, 'rewards/near_duplicate_penalty/std': 0.3419327139854431, 'rewards/opening_diversity/mean': 0.3902803361415863, 'rewards/opening_diversity/std': 0.39332956075668335, 'rewards/distinct_2/mean': 0.2485092431306839, 'rewards/distinct_2/std': 0.4601607620716095, 'rewards/total_composite/mean': 0.09274835884571075, 'rewards/total_composite/std': 0.05764620378613472, 'reward': 0.09274835884571075, 'reward_std': 0.05764620378613472, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.030050111934542656, 'sampling/sampling_logp_difference/max': 1.4638023376464844, 'sampling/importance_sampling_ratio/min': 0.23135492205619812, 'sampling/importance_sampling_ratio/mean': 1.005531907081604, 'sampling/importance_sampling_ratio/max': 1.9063942432403564, 'entropy': 0.19107347913086414, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.017522387322969735, 'clip_ratio/high_max': 0.017522387322969735, 'clip_ratio/region_mean': 0.017522387322969735, 'reward_total_mean': 0.09274835884571075, 'reward_meter_mean': 0.7749155759811401, 'reward_meter_std': 0.36531969904899597, 'reward_count_adherence_mean': 0.6607142686843872, 'reward_count_adherence_std': 0.36770108342170715, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9490639567375183, 'reward_lexical_plausibility_std': 0.0980229303240776, 'reward_repeat_penalty_mean': 0.21573305130004883, 'reward_repeat_penalty_std': 0.4061233401298523, 'reward_near_duplicate_penalty_mean': 0.4996299147605896, 'reward_near_duplicate_penalty_std': 0.3419327139854431, 'reward_opening_diversity_mean': 0.3902803361415863, 'reward_opening_diversity_std': 0.39332956075668335, 'reward_distinct_2_mean': 0.2485092431306839, 'reward_distinct_2_std': 0.4601607620716095, 'reward_judge_quality_mean': 0.125, 'reward_judge_quality_std': 0.046291008591651917, 'reward_total_composite_mean': 0.09274835884571075, 'reward_total_composite_std': 0.05764620378613472, 'epoch': 0.02}

 18%|█▊        | 599/3300 [2:20:57<9:50:35, 13.12s/it]INFO 04-12 13:07:09 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 600/3300 [2:21:11<10:07:51, 13.51s/it]2026-04-12 13:07:23,998 | INFO | train_grpo_train | metrics_logged mode=train step=600

                                                       
{'loss': -0.1437, 'grad_norm': 2.395522117614746, 'learning_rate': 8.184848484848486e-06, 'num_tokens': 1208870.0, 'completions/mean_length': 248.125, 'completions/min_length': 87.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 89.80000305175781, 'completions/min_terminated_length': 87.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.7769525051116943, 'rewards/meter/std': 0.2592514157295227, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.18898223340511322, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.901878833770752, 'rewards/lexical_plausibility/std': 0.14493505656719208, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.16690459847450256, 'rewards/repeat_penalty/mean': 0.7106788754463196, 'rewards/repeat_penalty/std': 0.3217860162258148, 'rewards/near_duplicate_penalty/mean': 0.9214297533035278, 'rewards/near_duplicate_penalty/std': 0.09038443118333817, 'rewards/opening_diversity/mean': 0.8415178656578064, 'rewards/opening_diversity/std': 0.2187955528497696, 'rewards/distinct_2/mean': 0.8476666212081909, 'rewards/distinct_2/std': 0.18866392970085144, 'rewards/total_composite/mean': 0.20124664902687073, 'rewards/total_composite/std': 0.17528368532657623, 'reward': 0.20124664902687073, 'reward_std': 0.17528368532657623, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13525071740150452, 'sampling/sampling_logp_difference/max': 1.5899407863616943, 'sampling/importance_sampling_ratio/min': 0.20393767952919006, 'sampling/importance_sampling_ratio/mean': 1.0231225490570068, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5701515153050423, 'clip_ratio/low_mean': 0.007102272938936949, 'clip_ratio/low_min': 0.007102272938936949, 'clip_ratio/high_mean': 0.0804181657731533, 'clip_ratio/high_max': 0.0804181657731533, 'clip_ratio/region_mean': 0.08752043871209025, 'reward_total_mean': 0.20124664902687073, 'reward_meter_mean': 0.7769525051116943, 'reward_meter_std': 0.2592514157295227, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.18898223340511322, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.901878833770752, 'reward_lexical_plausibility_std': 0.14493505656719208, 'reward_repeat_penalty_mean': 0.7106788754463196, 'reward_repeat_penalty_std': 0.3217860162258148, 'reward_near_duplicate_penalty_mean': 0.9214297533035278, 'reward_near_duplicate_penalty_std': 0.09038443118333817, 'reward_opening_diversity_mean': 0.8415178656578064, 'reward_opening_diversity_std': 0.2187955528497696, 'reward_distinct_2_mean': 0.8476666212081909, 'reward_distinct_2_std': 0.18866392970085144, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.16690459847450256, 'reward_total_composite_mean': 0.20124664902687073, 'reward_total_composite_std': 0.17528368532657623, 'epoch': 0.02}

 18%|█▊        | 600/3300 [2:21:11<10:07:51, 13.51s/it]INFO 04-12 13:07:24 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


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

 15%|█▌        | 2/13 [00:11<01:02,  5.72s/it]

 23%|██▎       | 3/13 [00:23<01:24,  8.46s/it]

 31%|███       | 4/13 [00:36<01:29,  9.91s/it]

 38%|███▊      | 5/13 [00:47<01:24, 10.58s/it]

 46%|████▌     | 6/13 [00:59<01:16, 10.91s/it]

 54%|█████▍    | 7/13 [01:08<01:01, 10.25s/it]

 62%|██████▏   | 8/13 [01:21<00:55, 11.07s/it]

 69%|██████▉   | 9/13 [01:34<00:47, 11.76s/it]

 77%|███████▋  | 10/13 [01:46<00:35, 11.87s/it]

 85%|████████▍ | 11/13 [01:56<00:22, 11.38s/it]

 92%|█████████▏| 12/13 [02:10<00:11, 11.94s/it]

100%|██████████| 13/13 [02:22<00:00, 12.06s/it]2026-04-12 13:09:58,350 | INFO | train_grpo_train | metrics_logged mode=eval step=600

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 154.3055, 'eval_samples_per_second': 0.674, 'eval_steps_per_second': 0.084, 'eval_num_tokens': 1208870.0, 'eval_completions/mean_length': 240.84615384615384, 'eval_completions/min_length': 49.76923076923077, 'eval_completions/max_length': 482.15384615384613, 'eval_completions/clipped_ratio': 0.23076923076923078, 'eval_completions/mean_terminated_length': 161.69963836669922, 'eval_completions/min_terminated_length': 49.76923076923077, 'eval_completions/max_terminated_length': 322.7692307692308, 'eval_rewards/meter/mean': 0.5123148812697484, 'eval_rewards/meter/std': 0.38864422073731053, 'eval_rewards/count_adherence/mean': 0.8358814578789932, 'eval_rewards/count_adherence/std': 0.24184314992565376, 'eval_rewards/arabic_clean/mean': 0.7788461538461539, 'eval_rewards/arabic_clean/std': 0.4062624206909767, 'eval_rewards/lexical_plausibility/mean': 0.9462600946426392, 'eval_rewards/lexical_plausibility/std': 0.09318991621526387, 'eval_rewards/judge_quality/mean': 0.2827884600712703, 'eval_rewards/judge_quality/std': 0.18655071120995742, 'eval_rewards/repeat_penalty/mean': 0.7113298040169936, 'eval_rewards/repeat_penalty/std': 0.35973638869248903, 'eval_rewards/near_duplicate_penalty/mean': 0.9038105607032776, 'eval_rewards/near_duplicate_penalty/std': 0.13724762401901758, 'eval_rewards/opening_diversity/mean': 0.9311670752672049, 'eval_rewards/opening_diversity/std': 0.12307891335624915, 'eval_rewards/distinct_2/mean': 0.7650806307792664, 'eval_rewards/distinct_2/std': 0.3487296878145291, 'eval_rewards/total_composite/mean': 0.13370628081835234, 'eval_rewards/total_composite/std': 0.1360558053621879, 'eval_reward': 0.13370628081835234, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.06334158462973741, 'eval_sampling/sampling_logp_difference/max': 1.0068141497098482, 'eval_sampling/importance_sampling_ratio/min': 0.36993258733015794, 'eval_sampling/importance_sampling_ratio/mean': 1.0165570974349976, 'eval_sampling/importance_sampling_ratio/max': 1.5171292561751146, 'eval_entropy': 0.7645946099207952, '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': 0.13370628081835234, 'eval_reward_meter_mean': 0.5123148812697484, 'eval_reward_meter_std': 0.38864422073731053, 'eval_reward_count_adherence_mean': 0.8358814578789932, 'eval_reward_count_adherence_std': 0.24184314992565376, 'eval_reward_arabic_clean_mean': 0.7788461538461539, 'eval_reward_arabic_clean_std': 0.4062624206909767, 'eval_reward_lexical_plausibility_mean': 0.9462600946426392, 'eval_reward_lexical_plausibility_std': 0.09318991621526387, 'eval_reward_repeat_penalty_mean': 0.7113298040169936, 'eval_reward_repeat_penalty_std': 0.35973638869248903, 'eval_reward_near_duplicate_penalty_mean': 0.9038105607032776, 'eval_reward_near_duplicate_penalty_std': 0.13724762401901758, 'eval_reward_opening_diversity_mean': 0.9311670752672049, 'eval_reward_opening_diversity_std': 0.12307891335624915, 'eval_reward_distinct_2_mean': 0.7650806307792664, 'eval_reward_distinct_2_std': 0.3487296878145291, 'eval_reward_judge_quality_mean': 0.2827884600712703, 'eval_reward_judge_quality_std': 0.18655071120995742, 'eval_reward_total_composite_mean': 0.13370628081835234, 'eval_reward_total_composite_std': 0.1360558053621879, 'epoch': 0.02}

 18%|█▊        | 600/3300 [2:23:46<10:07:51, 13.51s/it]

100%|██████████| 13/13 [02:22<00:00, 12.06s/it]

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

 18%|█▊        | 601/3300 [2:24:02<45:33:03, 60.76s/it]2026-04-12 13:10:15,005 | INFO | train_grpo_train | metrics_logged mode=train step=601

                                                       
{'loss': -0.0498, 'grad_norm': 1.9377853870391846, 'learning_rate': 8.181818181818183e-06, 'num_tokens': 1210352.0, 'completions/mean_length': 218.25, 'completions/min_length': 37.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 42.0, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.2855945825576782, 'rewards/meter/std': 0.35705050826072693, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.890587568283081, 'rewards/lexical_plausibility/std': 0.15190117061138153, 'rewards/judge_quality/mean': 0.33375000953674316, 'rewards/judge_quality/std': 0.29587340354919434, 'rewards/repeat_penalty/mean': 0.8381612300872803, 'rewards/repeat_penalty/std': 0.19212819635868073, 'rewards/near_duplicate_penalty/mean': 0.9620789289474487, 'rewards/near_duplicate_penalty/std': 0.052541568875312805, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9505941271781921, 'rewards/distinct_2/std': 0.10680097341537476, 'rewards/total_composite/mean': 0.10880088061094284, 'rewards/total_composite/std': 0.1451246440410614, 'reward': 0.10880088061094284, 'reward_std': 0.1451246440410614, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12478195130825043, 'sampling/sampling_logp_difference/max': 0.9111672639846802, 'sampling/importance_sampling_ratio/min': 0.40205466747283936, 'sampling/importance_sampling_ratio/mean': 1.024455189704895, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7010936513543129, 'clip_ratio/low_mean': 0.0299990214407444, 'clip_ratio/low_min': 0.0299990214407444, 'clip_ratio/high_mean': 0.02738095261156559, 'clip_ratio/high_max': 0.02738095261156559, 'clip_ratio/region_mean': 0.05737997405230999, 'reward_total_mean': 0.10880088061094284, 'reward_meter_mean': 0.2855945825576782, 'reward_meter_std': 0.35705050826072693, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.890587568283081, 'reward_lexical_plausibility_std': 0.15190117061138153, 'reward_repeat_penalty_mean': 0.8381612300872803, 'reward_repeat_penalty_std': 0.19212819635868073, 'reward_near_duplicate_penalty_mean': 0.9620789289474487, 'reward_near_duplicate_penalty_std': 0.052541568875312805, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9505941271781921, 'reward_distinct_2_std': 0.10680097341537476, 'reward_judge_quality_mean': 0.33375000953674316, 'reward_judge_quality_std': 0.29587340354919434, 'reward_total_composite_mean': 0.10880088061094284, 'reward_total_composite_std': 0.1451246440410614, 'epoch': 0.02}

 18%|█▊        | 601/3300 [2:24:02<45:33:03, 60.76s/it]INFO 04-12 13:10:15 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 602/3300 [2:24:11<33:50:33, 45.16s/it]2026-04-12 13:10:23,763 | INFO | train_grpo_train | metrics_logged mode=train step=602

                                                       
{'loss': 0.0884, 'grad_norm': 16.810245513916016, 'learning_rate': 8.17878787878788e-06, 'num_tokens': 1211849.0, 'completions/mean_length': 32.125, 'completions/min_length': 29.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.125, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.626380443572998, 'rewards/meter/std': 0.3864566683769226, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.71875, 'rewards/judge_quality/std': 0.2794605791568756, 'rewards/repeat_penalty/mean': 0.949288010597229, 'rewards/repeat_penalty/std': 0.04986143112182617, 'rewards/near_duplicate_penalty/mean': 0.9612667560577393, 'rewards/near_duplicate_penalty/std': 0.02451680414378643, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9871795177459717, 'rewards/distinct_2/std': 0.03626188635826111, 'rewards/total_composite/mean': 0.4113466739654541, 'rewards/total_composite/std': 0.3315257132053375, 'reward': 0.4113466739654541, 'reward_std': 0.3315257132053375, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1796492040157318, 'sampling/sampling_logp_difference/max': 1.7996973991394043, 'sampling/importance_sampling_ratio/min': 0.16534891724586487, 'sampling/importance_sampling_ratio/mean': 1.0236263275146484, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2363914623856544, 'clip_ratio/low_mean': 0.08908404689282179, 'clip_ratio/low_min': 0.08908404689282179, 'clip_ratio/high_mean': 0.05182291753590107, 'clip_ratio/high_max': 0.05182291753590107, 'clip_ratio/region_mean': 0.14090696442872286, 'reward_total_mean': 0.4113466739654541, 'reward_meter_mean': 0.626380443572998, 'reward_meter_std': 0.3864566683769226, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.949288010597229, 'reward_repeat_penalty_std': 0.04986143112182617, 'reward_near_duplicate_penalty_mean': 0.9612667560577393, 'reward_near_duplicate_penalty_std': 0.02451680414378643, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9871795177459717, 'reward_distinct_2_std': 0.03626188635826111, 'reward_judge_quality_mean': 0.71875, 'reward_judge_quality_std': 0.2794605791568756, 'reward_total_composite_mean': 0.4113466739654541, 'reward_total_composite_std': 0.3315257132053375, 'epoch': 0.02}

 18%|█▊        | 602/3300 [2:24:11<33:50:33, 45.16s/it]INFO 04-12 13:10:23 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 603/3300 [2:24:24<26:27:34, 35.32s/it]2026-04-12 13:10:36,122 | INFO | train_grpo_train | metrics_logged mode=train step=603

                                                       
{'loss': 0.0211, 'grad_norm': 3.903568744659424, 'learning_rate': 8.175757575757577e-06, 'num_tokens': 1215831.0, 'completions/mean_length': 284.75, 'completions/min_length': 259.0, 'completions/max_length': 351.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 284.75, 'completions/min_terminated_length': 259.0, 'completions/max_terminated_length': 351.0, 'rewards/meter/mean': 0.8659458160400391, 'rewards/meter/std': 0.21825405955314636, 'rewards/count_adherence/mean': 0.9807692170143127, 'rewards/count_adherence/std': 0.03560846298933029, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.20000000298023224, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.004345335997641087, 'rewards/repeat_penalty/std': 0.012290466576814651, 'rewards/near_duplicate_penalty/mean': 0.44887831807136536, 'rewards/near_duplicate_penalty/std': 0.1940241903066635, 'rewards/opening_diversity/mean': 0.49038463830947876, 'rewards/opening_diversity/std': 0.3376176357269287, 'rewards/distinct_2/mean': 0.008531960658729076, 'rewards/distinct_2/std': 0.024132030084729195, 'rewards/total_composite/mean': 0.1512950360774994, 'rewards/total_composite/std': 0.09996257722377777, 'reward': 0.1512950360774994, 'reward_std': 0.09996257722377777, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05523812770843506, 'sampling/sampling_logp_difference/max': 3.0551135540008545, 'sampling/importance_sampling_ratio/min': 0.04711737111210823, 'sampling/importance_sampling_ratio/mean': 1.0041542053222656, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.35286697559058666, 'clip_ratio/low_mean': 0.02745216037146747, 'clip_ratio/low_min': 0.02745216037146747, 'clip_ratio/high_mean': 0.01791114266961813, 'clip_ratio/high_max': 0.01791114266961813, 'clip_ratio/region_mean': 0.0453633030410856, 'reward_total_mean': 0.1512950360774994, 'reward_meter_mean': 0.8659458160400391, 'reward_meter_std': 0.21825405955314636, 'reward_count_adherence_mean': 0.9807692170143127, 'reward_count_adherence_std': 0.03560846298933029, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.004345335997641087, 'reward_repeat_penalty_std': 0.012290466576814651, 'reward_near_duplicate_penalty_mean': 0.44887831807136536, 'reward_near_duplicate_penalty_std': 0.1940241903066635, 'reward_opening_diversity_mean': 0.49038463830947876, 'reward_opening_diversity_std': 0.3376176357269287, 'reward_distinct_2_mean': 0.008531960658729076, 'reward_distinct_2_std': 0.024132030084729195, 'reward_judge_quality_mean': 0.20000000298023224, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.1512950360774994, 'reward_total_composite_std': 0.09996257722377777, 'epoch': 0.02}

 18%|█▊        | 603/3300 [2:24:24<26:27:34, 35.32s/it]INFO 04-12 13:10:36 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 604/3300 [2:24:37<21:31:35, 28.74s/it]2026-04-12 13:10:49,553 | INFO | train_grpo_train | metrics_logged mode=train step=604

                                                       
{'loss': -0.041, 'grad_norm': 4.592195987701416, 'learning_rate': 8.172727272727273e-06, 'num_tokens': 1217294.0, 'completions/mean_length': 100.875, 'completions/min_length': 37.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 42.142860412597656, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.46032580733299255, 'rewards/meter/std': 0.3890128433704376, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9694451689720154, 'rewards/lexical_plausibility/std': 0.08642213046550751, 'rewards/judge_quality/mean': 0.53125, 'rewards/judge_quality/std': 0.30600830912590027, 'rewards/repeat_penalty/mean': 0.8391133546829224, 'rewards/repeat_penalty/std': 0.30781739950180054, 'rewards/near_duplicate_penalty/mean': 0.9558348655700684, 'rewards/near_duplicate_penalty/std': 0.0789102241396904, 'rewards/opening_diversity/mean': 0.890625, 'rewards/opening_diversity/std': 0.2259652018547058, 'rewards/distinct_2/mean': 0.930601954460144, 'rewards/distinct_2/std': 0.14452484250068665, 'rewards/total_composite/mean': 0.31847354769706726, 'rewards/total_composite/std': 0.32631081342697144, 'reward': 0.31847354769706726, 'reward_std': 0.32631081342697144, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14562717080116272, 'sampling/sampling_logp_difference/max': 1.5322856903076172, 'sampling/importance_sampling_ratio/min': 0.21604129672050476, 'sampling/importance_sampling_ratio/mean': 0.9944778680801392, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8797452300786972, 'clip_ratio/low_mean': 0.04144576843827963, 'clip_ratio/low_min': 0.04144576843827963, 'clip_ratio/high_mean': 0.07772436086088419, 'clip_ratio/high_max': 0.07772436086088419, 'clip_ratio/region_mean': 0.11917012929916382, 'reward_total_mean': 0.31847354769706726, 'reward_meter_mean': 0.46032580733299255, 'reward_meter_std': 0.3890128433704376, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9694451689720154, 'reward_lexical_plausibility_std': 0.08642213046550751, 'reward_repeat_penalty_mean': 0.8391133546829224, 'reward_repeat_penalty_std': 0.30781739950180054, 'reward_near_duplicate_penalty_mean': 0.9558348655700684, 'reward_near_duplicate_penalty_std': 0.0789102241396904, 'reward_opening_diversity_mean': 0.890625, 'reward_opening_diversity_std': 0.2259652018547058, 'reward_distinct_2_mean': 0.930601954460144, 'reward_distinct_2_std': 0.14452484250068665, 'reward_judge_quality_mean': 0.53125, 'reward_judge_quality_std': 0.30600830912590027, 'reward_total_composite_mean': 0.31847354769706726, 'reward_total_composite_std': 0.32631081342697144, 'epoch': 0.02}

 18%|█▊        | 604/3300 [2:24:37<21:31:35, 28.74s/it]INFO 04-12 13:10:49 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 605/3300 [2:24:47<17:23:50, 23.24s/it]2026-04-12 13:10:59,923 | INFO | train_grpo_train | metrics_logged mode=train step=605

                                                       
{'loss': 0.0157, 'grad_norm': 6.320730209350586, 'learning_rate': 8.16969696969697e-06, 'num_tokens': 1219948.0, 'completions/mean_length': 117.75, 'completions/min_length': 90.0, 'completions/max_length': 137.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 117.75, 'completions/min_terminated_length': 90.0, 'completions/max_terminated_length': 137.0, 'rewards/meter/mean': 0.9773919582366943, 'rewards/meter/std': 0.013674728572368622, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.24494178593158722, 'rewards/repeat_penalty/std': 0.1995166391134262, 'rewards/near_duplicate_penalty/mean': 0.7836527824401855, 'rewards/near_duplicate_penalty/std': 0.1670757383108139, 'rewards/opening_diversity/mean': 0.5625, 'rewards/opening_diversity/std': 0.2748376131057739, 'rewards/distinct_2/mean': 0.4683367908000946, 'rewards/distinct_2/std': 0.21342182159423828, 'rewards/total_composite/mean': 0.21540188789367676, 'rewards/total_composite/std': 0.10226933658123016, 'reward': 0.21540188789367676, 'reward_std': 0.10226932913064957, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10175961256027222, 'sampling/sampling_logp_difference/max': 1.4451415538787842, 'sampling/importance_sampling_ratio/min': 0.2357127070426941, 'sampling/importance_sampling_ratio/mean': 1.0097334384918213, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7508847042918205, 'clip_ratio/low_mean': 0.06030849181115627, 'clip_ratio/low_min': 0.06030849181115627, 'clip_ratio/high_mean': 0.046880092471838, 'clip_ratio/high_max': 0.046880092471838, 'clip_ratio/region_mean': 0.10718858428299427, 'reward_total_mean': 0.21540188789367676, 'reward_meter_mean': 0.9773919582366943, 'reward_meter_std': 0.013674728572368622, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.24494178593158722, 'reward_repeat_penalty_std': 0.1995166391134262, 'reward_near_duplicate_penalty_mean': 0.7836527824401855, 'reward_near_duplicate_penalty_std': 0.1670757383108139, 'reward_opening_diversity_mean': 0.5625, 'reward_opening_diversity_std': 0.2748376131057739, 'reward_distinct_2_mean': 0.4683367908000946, 'reward_distinct_2_std': 0.21342182159423828, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.21540188789367676, 'reward_total_composite_std': 0.10226933658123016, 'epoch': 0.02}

 18%|█▊        | 605/3300 [2:24:47<17:23:50, 23.24s/it]INFO 04-12 13:11:00 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 606/3300 [2:24:56<14:05:11, 18.82s/it]2026-04-12 13:11:08,442 | INFO | train_grpo_train | metrics_logged mode=train step=606

                                                       
{'loss': -0.0129, 'grad_norm': 19.67855453491211, 'learning_rate': 8.166666666666668e-06, 'num_tokens': 1221539.0, 'completions/mean_length': 34.875, 'completions/min_length': 29.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.875, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.7507923245429993, 'rewards/meter/std': 0.2904006540775299, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7900000214576721, 'rewards/judge_quality/std': 0.24219238758087158, 'rewards/repeat_penalty/mean': 0.9974340200424194, 'rewards/repeat_penalty/std': 0.007257687393575907, 'rewards/near_duplicate_penalty/mean': 0.9974340200424194, 'rewards/near_duplicate_penalty/std': 0.007257687393575907, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5838419198989868, 'rewards/total_composite/std': 0.30834606289863586, 'reward': 0.5838419198989868, 'reward_std': 0.30834606289863586, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14887173473834991, 'sampling/sampling_logp_difference/max': 1.9746841192245483, 'sampling/importance_sampling_ratio/min': 0.13880515098571777, 'sampling/importance_sampling_ratio/mean': 0.9879239797592163, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.658489890396595, 'clip_ratio/low_mean': 0.06835006643086672, 'clip_ratio/low_min': 0.06835006643086672, 'clip_ratio/high_mean': 0.0649535283446312, 'clip_ratio/high_max': 0.0649535283446312, 'clip_ratio/region_mean': 0.1333035947754979, 'reward_total_mean': 0.5838419198989868, 'reward_meter_mean': 0.7507923245429993, 'reward_meter_std': 0.2904006540775299, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9974340200424194, 'reward_repeat_penalty_std': 0.007257687393575907, 'reward_near_duplicate_penalty_mean': 0.9974340200424194, 'reward_near_duplicate_penalty_std': 0.007257687393575907, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7900000214576721, 'reward_judge_quality_std': 0.24219238758087158, 'reward_total_composite_mean': 0.5838419198989868, 'reward_total_composite_std': 0.30834606289863586, 'epoch': 0.02}

 18%|█▊        | 606/3300 [2:24:56<14:05:11, 18.82s/it]INFO 04-12 13:11:08 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 607/3300 [2:25:06<12:08:04, 16.22s/it]2026-04-12 13:11:18,592 | INFO | train_grpo_train | metrics_logged mode=train step=607

                                                       
{'loss': -0.0017, 'grad_norm': 11.003536224365234, 'learning_rate': 8.163636363636365e-06, 'num_tokens': 1223422.0, 'completions/mean_length': 68.375, 'completions/min_length': 63.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.375, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.9552795886993408, 'rewards/meter/std': 0.050910934805870056, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4962500035762787, 'rewards/judge_quality/std': 0.17476005852222443, 'rewards/repeat_penalty/mean': 0.7330594062805176, 'rewards/repeat_penalty/std': 0.20595264434814453, 'rewards/near_duplicate_penalty/mean': 0.9217890501022339, 'rewards/near_duplicate_penalty/std': 0.058528222143650055, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.8583263158798218, 'rewards/distinct_2/std': 0.09201016277074814, 'rewards/total_composite/mean': 0.47105875611305237, 'rewards/total_composite/std': 0.15274225175380707, 'reward': 0.47105875611305237, 'reward_std': 0.15274225175380707, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14251470565795898, 'sampling/sampling_logp_difference/max': 1.588109016418457, 'sampling/importance_sampling_ratio/min': 0.20431159436702728, 'sampling/importance_sampling_ratio/mean': 1.0019904375076294, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7803902328014374, 'clip_ratio/low_mean': 0.08321047015488148, 'clip_ratio/low_min': 0.08321047015488148, 'clip_ratio/high_mean': 0.02678571455180645, 'clip_ratio/high_max': 0.02678571455180645, 'clip_ratio/region_mean': 0.10999618470668793, 'reward_total_mean': 0.47105875611305237, 'reward_meter_mean': 0.9552795886993408, 'reward_meter_std': 0.050910934805870056, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7330594062805176, 'reward_repeat_penalty_std': 0.20595264434814453, 'reward_near_duplicate_penalty_mean': 0.9217890501022339, 'reward_near_duplicate_penalty_std': 0.058528222143650055, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.8583263158798218, 'reward_distinct_2_std': 0.09201016277074814, 'reward_judge_quality_mean': 0.4962500035762787, 'reward_judge_quality_std': 0.17476005852222443, 'reward_total_composite_mean': 0.47105875611305237, 'reward_total_composite_std': 0.15274225175380707, 'epoch': 0.02}

 18%|█▊        | 607/3300 [2:25:06<12:08:04, 16.22s/it]INFO 04-12 13:11:18 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 608/3300 [2:25:15<10:30:10, 14.05s/it]2026-04-12 13:11:27,562 | INFO | train_grpo_train | metrics_logged mode=train step=608

                                                       
{'loss': 0.0353, 'grad_norm': 12.383308410644531, 'learning_rate': 8.16060606060606e-06, 'num_tokens': 1225327.0, 'completions/mean_length': 53.125, 'completions/min_length': 46.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.125, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.7017092704772949, 'rewards/meter/std': 0.36034920811653137, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9866071343421936, 'rewards/lexical_plausibility/std': 0.03788072615861893, 'rewards/judge_quality/mean': 0.4625000059604645, 'rewards/judge_quality/std': 0.12464234977960587, 'rewards/repeat_penalty/mean': 0.9731595516204834, 'rewards/repeat_penalty/std': 0.05205292999744415, 'rewards/near_duplicate_penalty/mean': 0.9880070686340332, 'rewards/near_duplicate_penalty/std': 0.012963104993104935, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9846153855323792, 'rewards/distinct_2/std': 0.04351425915956497, 'rewards/total_composite/mean': 0.33394530415534973, 'rewards/total_composite/std': 0.20485422015190125, 'reward': 0.33394530415534973, 'reward_std': 0.20485422015190125, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17869995534420013, 'sampling/sampling_logp_difference/max': 1.3907296657562256, 'sampling/importance_sampling_ratio/min': 0.2751923203468323, 'sampling/importance_sampling_ratio/mean': 1.0368835926055908, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4442286118865013, 'clip_ratio/low_mean': 0.04861699044704437, 'clip_ratio/low_min': 0.04861699044704437, 'clip_ratio/high_mean': 0.1034470684826374, 'clip_ratio/high_max': 0.1034470684826374, 'clip_ratio/region_mean': 0.15206405892968178, 'reward_total_mean': 0.33394530415534973, 'reward_meter_mean': 0.7017092704772949, 'reward_meter_std': 0.36034920811653137, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9866071343421936, 'reward_lexical_plausibility_std': 0.03788072615861893, 'reward_repeat_penalty_mean': 0.9731595516204834, 'reward_repeat_penalty_std': 0.05205292999744415, 'reward_near_duplicate_penalty_mean': 0.9880070686340332, 'reward_near_duplicate_penalty_std': 0.012963104993104935, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9846153855323792, 'reward_distinct_2_std': 0.04351425915956497, 'reward_judge_quality_mean': 0.4625000059604645, 'reward_judge_quality_std': 0.12464234977960587, 'reward_total_composite_mean': 0.33394530415534973, 'reward_total_composite_std': 0.20485422015190125, 'epoch': 0.02}

 18%|█▊        | 608/3300 [2:25:15<10:30:10, 14.05s/it]INFO 04-12 13:11:27 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 609/3300 [2:25:25<9:30:36, 12.72s/it] 2026-04-12 13:11:37,197 | INFO | train_grpo_train | metrics_logged mode=train step=609

                                                      
{'loss': 0.0899, 'grad_norm': 16.349994659423828, 'learning_rate': 8.15757575757576e-06, 'num_tokens': 1227038.0, 'completions/mean_length': 53.875, 'completions/min_length': 47.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.875, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.7870069742202759, 'rewards/meter/std': 0.3712165653705597, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.41249996423721313, 'rewards/judge_quality/std': 0.14078859984874725, 'rewards/repeat_penalty/mean': 0.9397884011268616, 'rewards/repeat_penalty/std': 0.05675487220287323, 'rewards/near_duplicate_penalty/mean': 0.9605423212051392, 'rewards/near_duplicate_penalty/std': 0.03168892115354538, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9780626893043518, 'rewards/distinct_2/std': 0.04296828806400299, 'rewards/total_composite/mean': 0.3579288721084595, 'rewards/total_composite/std': 0.20732474327087402, 'reward': 0.3579288721084595, 'reward_std': 0.20732474327087402, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14755448698997498, 'sampling/sampling_logp_difference/max': 2.3139448165893555, 'sampling/importance_sampling_ratio/min': 0.09887045621871948, 'sampling/importance_sampling_ratio/mean': 1.017095923423767, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0073984190821648, 'clip_ratio/low_mean': 0.05191622208803892, 'clip_ratio/low_min': 0.05191622208803892, 'clip_ratio/high_mean': 0.11532354354858398, 'clip_ratio/high_max': 0.11532354354858398, 'clip_ratio/region_mean': 0.1672397656366229, 'reward_total_mean': 0.3579288721084595, 'reward_meter_mean': 0.7870069742202759, 'reward_meter_std': 0.3712165653705597, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9397884011268616, 'reward_repeat_penalty_std': 0.05675487220287323, 'reward_near_duplicate_penalty_mean': 0.9605423212051392, 'reward_near_duplicate_penalty_std': 0.03168892115354538, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9780626893043518, 'reward_distinct_2_std': 0.04296828806400299, 'reward_judge_quality_mean': 0.41249996423721313, 'reward_judge_quality_std': 0.14078859984874725, 'reward_total_composite_mean': 0.3579288721084595, 'reward_total_composite_std': 0.20732474327087402, 'epoch': 0.02}

 18%|█▊        | 609/3300 [2:25:25<9:30:36, 12.72s/it]INFO 04-12 13:11:37 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 610/3300 [2:25:39<9:57:38, 13.33s/it]2026-04-12 13:11:51,946 | INFO | train_grpo_train | metrics_logged mode=train step=610

                                                      
{'loss': -0.0639, 'grad_norm': 5.973407745361328, 'learning_rate': 8.154545454545455e-06, 'num_tokens': 1228932.0, 'completions/mean_length': 113.75, 'completions/min_length': 52.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 56.857147216796875, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.7780847549438477, 'rewards/meter/std': 0.38483375310897827, '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/lexical_plausibility/mean': 0.9737052917480469, 'rewards/lexical_plausibility/std': 0.07437256723642349, 'rewards/judge_quality/mean': 0.39375001192092896, 'rewards/judge_quality/std': 0.18407586216926575, 'rewards/repeat_penalty/mean': 0.9340187311172485, 'rewards/repeat_penalty/std': 0.09607148170471191, 'rewards/near_duplicate_penalty/mean': 0.9705223441123962, 'rewards/near_duplicate_penalty/std': 0.032134246081113815, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9606755375862122, 'rewards/distinct_2/std': 0.07428864389657974, 'rewards/total_composite/mean': 0.30759817361831665, 'rewards/total_composite/std': 0.23372232913970947, 'reward': 0.30759817361831665, 'reward_std': 0.23372232913970947, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14644905924797058, 'sampling/sampling_logp_difference/max': 1.707603931427002, 'sampling/importance_sampling_ratio/min': 0.18129968643188477, 'sampling/importance_sampling_ratio/mean': 1.022405743598938, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8063682839274406, 'clip_ratio/low_mean': 0.03767339326441288, 'clip_ratio/low_min': 0.03767339326441288, 'clip_ratio/high_mean': 0.07242566626518965, 'clip_ratio/high_max': 0.07242566626518965, 'clip_ratio/region_mean': 0.11009905952960253, 'reward_total_mean': 0.30759817361831665, 'reward_meter_mean': 0.7780847549438477, 'reward_meter_std': 0.38483375310897827, '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_lexical_plausibility_mean': 0.9737052917480469, 'reward_lexical_plausibility_std': 0.07437256723642349, 'reward_repeat_penalty_mean': 0.9340187311172485, 'reward_repeat_penalty_std': 0.09607148170471191, 'reward_near_duplicate_penalty_mean': 0.9705223441123962, 'reward_near_duplicate_penalty_std': 0.032134246081113815, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9606755375862122, 'reward_distinct_2_std': 0.07428864389657974, 'reward_judge_quality_mean': 0.39375001192092896, 'reward_judge_quality_std': 0.18407586216926575, 'reward_total_composite_mean': 0.30759817361831665, 'reward_total_composite_std': 0.23372232913970947, 'epoch': 0.02}

 18%|█▊        | 610/3300 [2:25:39<9:57:38, 13.33s/it]INFO 04-12 13:11:52 [block_pool.py:378] Successfully reset prefix cache

 19%|█▊        | 611/3300 [2:25:53<10:07:19, 13.55s/it]2026-04-12 13:12:06,009 | INFO | train_grpo_train | metrics_logged mode=train step=611

                                                       
{'loss': -0.028, 'grad_norm': 5.934566497802734, 'learning_rate': 8.151515151515152e-06, 'num_tokens': 1230642.0, 'completions/mean_length': 109.75, 'completions/min_length': 47.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 52.28571701049805, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.42813026905059814, 'rewards/meter/std': 0.3249087631702423, '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/lexical_plausibility/mean': 0.9751070737838745, 'rewards/lexical_plausibility/std': 0.07040776312351227, 'rewards/judge_quality/mean': 0.5012500286102295, 'rewards/judge_quality/std': 0.36286312341690063, 'rewards/repeat_penalty/mean': 0.9826894402503967, 'rewards/repeat_penalty/std': 0.027173364534974098, 'rewards/near_duplicate_penalty/mean': 0.9913424253463745, 'rewards/near_duplicate_penalty/std': 0.010528694838285446, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9912587404251099, 'rewards/distinct_2/std': 0.024724015966057777, 'rewards/total_composite/mean': 0.19174829125404358, 'rewards/total_composite/std': 0.21112436056137085, 'reward': 0.19174829125404358, 'reward_std': 0.21112436056137085, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15664783120155334, 'sampling/sampling_logp_difference/max': 2.4705190658569336, 'sampling/importance_sampling_ratio/min': 0.0845409631729126, 'sampling/importance_sampling_ratio/mean': 1.0318156480789185, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9008237831294537, 'clip_ratio/low_mean': 0.08856004476547241, 'clip_ratio/low_min': 0.08856004476547241, 'clip_ratio/high_mean': 0.054008278995752335, 'clip_ratio/high_max': 0.054008278995752335, 'clip_ratio/region_mean': 0.14256832376122475, 'reward_total_mean': 0.19174829125404358, 'reward_meter_mean': 0.42813026905059814, 'reward_meter_std': 0.3249087631702423, '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_lexical_plausibility_mean': 0.9751070737838745, 'reward_lexical_plausibility_std': 0.07040776312351227, 'reward_repeat_penalty_mean': 0.9826894402503967, 'reward_repeat_penalty_std': 0.027173364534974098, 'reward_near_duplicate_penalty_mean': 0.9913424253463745, 'reward_near_duplicate_penalty_std': 0.010528694838285446, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9912587404251099, 'reward_distinct_2_std': 0.024724015966057777, 'reward_judge_quality_mean': 0.5012500286102295, 'reward_judge_quality_std': 0.36286312341690063, 'reward_total_composite_mean': 0.19174829125404358, 'reward_total_composite_std': 0.21112436056137085, 'epoch': 0.02}

 19%|█▊        | 611/3300 [2:25:53<10:07:19, 13.55s/it]INFO 04-12 13:12:06 [block_pool.py:378] Successfully reset prefix cache

 19%|█▊        | 612/3300 [2:26:02<8:59:24, 12.04s/it] 2026-04-12 13:12:14,525 | INFO | train_grpo_train | metrics_logged mode=train step=612

                                                      
{'loss': 0.0116, 'grad_norm': 21.2740478515625, 'learning_rate': 8.14848484848485e-06, 'num_tokens': 1231954.0, 'completions/mean_length': 21.0, 'completions/min_length': 17.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.0, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.961848258972168, 'rewards/meter/std': 0.053047578781843185, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4300000071525574, 'rewards/judge_quality/std': 0.32262319326400757, 'rewards/repeat_penalty/mean': 0.46212121844291687, 'rewards/repeat_penalty/std': 0.3830924332141876, 'rewards/near_duplicate_penalty/mean': 0.8724478483200073, 'rewards/near_duplicate_penalty/std': 0.16538074612617493, 'rewards/opening_diversity/mean': 0.46875, 'rewards/opening_diversity/std': 0.38816189765930176, 'rewards/distinct_2/mean': 0.884615421295166, 'rewards/distinct_2/std': 0.15924589335918427, 'rewards/total_composite/mean': 0.4156397879123688, 'rewards/total_composite/std': 0.30976879596710205, 'reward': 0.4156397879123688, 'reward_std': 0.30976879596710205, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11739847809076309, 'sampling/sampling_logp_difference/max': 1.1665740013122559, 'sampling/importance_sampling_ratio/min': 0.3114320635795593, 'sampling/importance_sampling_ratio/mean': 1.0275083780288696, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7884392067790031, 'clip_ratio/low_mean': 0.0880969543941319, 'clip_ratio/low_min': 0.0880969543941319, 'clip_ratio/high_mean': 0.004807692486792803, 'clip_ratio/high_max': 0.004807692486792803, 'clip_ratio/region_mean': 0.0929046468809247, 'reward_total_mean': 0.4156397879123688, 'reward_meter_mean': 0.961848258972168, 'reward_meter_std': 0.053047578781843185, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.46212121844291687, 'reward_repeat_penalty_std': 0.3830924332141876, 'reward_near_duplicate_penalty_mean': 0.8724478483200073, 'reward_near_duplicate_penalty_std': 0.16538074612617493, 'reward_opening_diversity_mean': 0.46875, 'reward_opening_diversity_std': 0.38816189765930176, 'reward_distinct_2_mean': 0.884615421295166, 'reward_distinct_2_std': 0.15924589335918427, 'reward_judge_quality_mean': 0.4300000071525574, 'reward_judge_quality_std': 0.32262319326400757, 'reward_total_composite_mean': 0.4156397879123688, 'reward_total_composite_std': 0.30976879596710205, 'epoch': 0.02}

 19%|█▊        | 612/3300 [2:26:02<8:59:24, 12.04s/it]INFO 04-12 13:12:14 [block_pool.py:378] Successfully reset prefix cache

 19%|█▊        | 613/3300 [2:26:10<8:05:46, 10.85s/it]2026-04-12 13:12:22,589 | INFO | train_grpo_train | metrics_logged mode=train step=613

                                                      
{'loss': -0.0054, 'grad_norm': 19.19696617126465, 'learning_rate': 8.145454545454547e-06, 'num_tokens': 1233401.0, 'completions/mean_length': 33.875, 'completions/min_length': 30.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.875, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.49567335844039917, 'rewards/meter/std': 0.407177209854126, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.7437499761581421, 'rewards/judge_quality/std': 0.2432481348514557, 'rewards/repeat_penalty/mean': 0.9988636374473572, 'rewards/repeat_penalty/std': 0.003214118769392371, 'rewards/near_duplicate_penalty/mean': 0.9988636374473572, 'rewards/near_duplicate_penalty/std': 0.003214118769392371, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.38026097416877747, 'rewards/total_composite/std': 0.3626217246055603, 'reward': 0.38026097416877747, 'reward_std': 0.3626217246055603, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16059303283691406, 'sampling/sampling_logp_difference/max': 1.1132054328918457, 'sampling/importance_sampling_ratio/min': 0.3285042643547058, 'sampling/importance_sampling_ratio/mean': 1.009480595588684, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8764260411262512, 'clip_ratio/low_mean': 0.06027056369930506, 'clip_ratio/low_min': 0.06027056369930506, 'clip_ratio/high_mean': 0.06838949210941792, 'clip_ratio/high_max': 0.06838949210941792, 'clip_ratio/region_mean': 0.12866005580872297, 'reward_total_mean': 0.38026097416877747, 'reward_meter_mean': 0.49567335844039917, 'reward_meter_std': 0.407177209854126, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.9988636374473572, 'reward_repeat_penalty_std': 0.003214118769392371, 'reward_near_duplicate_penalty_mean': 0.9988636374473572, 'reward_near_duplicate_penalty_std': 0.003214118769392371, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7437499761581421, 'reward_judge_quality_std': 0.2432481348514557, 'reward_total_composite_mean': 0.38026097416877747, 'reward_total_composite_std': 0.3626217246055603, 'epoch': 0.02}

 19%|█▊        | 613/3300 [2:26:10<8:05:46, 10.85s/it]INFO 04-12 13:12:22 [block_pool.py:378] Successfully reset prefix cache

 19%|█▊        | 614/3300 [2:26:20<7:52:53, 10.56s/it]2026-04-12 13:12:32,491 | INFO | train_grpo_train | metrics_logged mode=train step=614

                                                      
{'loss': 0.0478, 'grad_norm': 7.49420690536499, 'learning_rate': 8.142424242424242e-06, 'num_tokens': 1236292.0, 'completions/mean_length': 168.375, 'completions/min_length': 148.0, 'completions/max_length': 198.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 168.375, 'completions/min_terminated_length': 148.0, 'completions/max_terminated_length': 198.0, 'rewards/meter/mean': 0.9708923101425171, 'rewards/meter/std': 0.04286763817071915, 'rewards/count_adherence/mean': 0.890625, 'rewards/count_adherence/std': 0.04419417306780815, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.0353553369641304, 'rewards/repeat_penalty/mean': 0.6537837982177734, 'rewards/repeat_penalty/std': 0.27695295214653015, 'rewards/near_duplicate_penalty/mean': 0.9114950895309448, 'rewards/near_duplicate_penalty/std': 0.08545321226119995, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.7120804786682129, 'rewards/distinct_2/std': 0.2674011290073395, 'rewards/total_composite/mean': 0.19515065848827362, 'rewards/total_composite/std': 0.16530510783195496, 'reward': 0.19515065848827362, 'reward_std': 0.16530510783195496, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12216014415025711, 'sampling/sampling_logp_difference/max': 1.8836517333984375, 'sampling/importance_sampling_ratio/min': 0.15203391015529633, 'sampling/importance_sampling_ratio/mean': 1.0056089162826538, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8603750579059124, 'clip_ratio/low_mean': 0.03700372390449047, 'clip_ratio/low_min': 0.03700372390449047, 'clip_ratio/high_mean': 0.09614876471459866, 'clip_ratio/high_max': 0.09614876471459866, 'clip_ratio/region_mean': 0.13315248861908913, 'reward_total_mean': 0.19515065848827362, 'reward_meter_mean': 0.9708923101425171, 'reward_meter_std': 0.04286763817071915, 'reward_count_adherence_mean': 0.890625, 'reward_count_adherence_std': 0.04419417306780815, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6537837982177734, 'reward_repeat_penalty_std': 0.27695295214653015, 'reward_near_duplicate_penalty_mean': 0.9114950895309448, 'reward_near_duplicate_penalty_std': 0.08545321226119995, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.7120804786682129, 'reward_distinct_2_std': 0.2674011290073395, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.0353553369641304, 'reward_total_composite_mean': 0.19515065848827362, 'reward_total_composite_std': 0.16530510783195496, 'epoch': 0.02}

 19%|█▊        | 614/3300 [2:26:20<7:52:53, 10.56s/it]INFO 04-12 13:12:32 [block_pool.py:378] Successfully reset prefix cache

 19%|█▊        | 615/3300 [2:26:27<7:10:19,  9.62s/it]2026-04-12 13:12:39,896 | INFO | train_grpo_train | metrics_logged mode=train step=615

                                                      
{'loss': 0.0385, 'grad_norm': 20.74620246887207, 'learning_rate': 8.139393939393941e-06, 'num_tokens': 1237708.0, 'completions/mean_length': 22.0, 'completions/min_length': 19.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.0, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.6093144416809082, 'rewards/meter/std': 0.3814626634120941, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7649999856948853, 'rewards/judge_quality/std': 0.2979933023452759, 'rewards/repeat_penalty/mean': 0.7414772510528564, 'rewards/repeat_penalty/std': 0.02410591021180153, 'rewards/near_duplicate_penalty/mean': 0.9886363744735718, 'rewards/near_duplicate_penalty/std': 0.03214120864868164, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4846035838127136, 'rewards/total_composite/std': 0.35006004571914673, 'reward': 0.4846035838127136, 'reward_std': 0.35006001591682434, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17349903285503387, 'sampling/sampling_logp_difference/max': 1.5350637435913086, 'sampling/importance_sampling_ratio/min': 0.35379427671432495, 'sampling/importance_sampling_ratio/mean': 1.0315041542053223, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9633065909147263, 'clip_ratio/low_mean': 0.12268170714378357, 'clip_ratio/low_min': 0.12268170714378357, 'clip_ratio/high_mean': 0.038784585893154144, 'clip_ratio/high_max': 0.038784585893154144, 'clip_ratio/region_mean': 0.1614662930369377, 'reward_total_mean': 0.4846035838127136, 'reward_meter_mean': 0.6093144416809082, 'reward_meter_std': 0.3814626634120941, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7414772510528564, 'reward_repeat_penalty_std': 0.02410591021180153, 'reward_near_duplicate_penalty_mean': 0.9886363744735718, 'reward_near_duplicate_penalty_std': 0.03214120864868164, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7649999856948853, 'reward_judge_quality_std': 0.2979933023452759, 'reward_total_composite_mean': 0.4846035838127136, 'reward_total_composite_std': 0.35006004571914673, 'epoch': 0.02}

 19%|█▊        | 615/3300 [2:26:27<7:10:19,  9.62s/it]INFO 04-12 13:12:40 [block_pool.py:378] Successfully reset prefix cache

 19%|█▊        | 616/3300 [2:26:36<6:52:24,  9.22s/it]2026-04-12 13:12:48,186 | INFO | train_grpo_train | metrics_logged mode=train step=616

                                                      
{'loss': 0.0166, 'grad_norm': 15.133147239685059, 'learning_rate': 8.136363636363637e-06, 'num_tokens': 1239337.0, 'completions/mean_length': 48.625, 'completions/min_length': 45.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.625, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.9023321866989136, 'rewards/meter/std': 0.1227162629365921, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9886363744735718, 'rewards/lexical_plausibility/std': 0.03214120864868164, 'rewards/judge_quality/mean': 0.7925000190734863, 'rewards/judge_quality/std': 0.21211522817611694, 'rewards/repeat_penalty/mean': 0.8279917240142822, 'rewards/repeat_penalty/std': 0.20133918523788452, 'rewards/near_duplicate_penalty/mean': 0.9511077404022217, 'rewards/near_duplicate_penalty/std': 0.04021043702960014, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9458041787147522, 'rewards/distinct_2/std': 0.08372856676578522, 'rewards/total_composite/mean': 0.706574022769928, 'rewards/total_composite/std': 0.19633522629737854, 'reward': 0.706574022769928, 'reward_std': 0.19633522629737854, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15388040244579315, 'sampling/sampling_logp_difference/max': 2.0758848190307617, 'sampling/importance_sampling_ratio/min': 0.1254453808069229, 'sampling/importance_sampling_ratio/mean': 0.9870394468307495, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9838912785053253, 'clip_ratio/low_mean': 0.06094952765852213, 'clip_ratio/low_min': 0.06094952765852213, 'clip_ratio/high_mean': 0.08726144209504128, 'clip_ratio/high_max': 0.08726144209504128, 'clip_ratio/region_mean': 0.1482109697535634, 'reward_total_mean': 0.706574022769928, 'reward_meter_mean': 0.9023321866989136, 'reward_meter_std': 0.1227162629365921, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9886363744735718, 'reward_lexical_plausibility_std': 0.03214120864868164, 'reward_repeat_penalty_mean': 0.8279917240142822, 'reward_repeat_penalty_std': 0.20133918523788452, 'reward_near_duplicate_penalty_mean': 0.9511077404022217, 'reward_near_duplicate_penalty_std': 0.04021043702960014, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9458041787147522, 'reward_distinct_2_std': 0.08372856676578522, 'reward_judge_quality_mean': 0.7925000190734863, 'reward_judge_quality_std': 0.21211522817611694, 'reward_total_composite_mean': 0.706574022769928, 'reward_total_composite_std': 0.19633522629737854, 'epoch': 0.02}

 19%|█▊        | 616/3300 [2:26:36<6:52:24,  9.22s/it]INFO 04-12 13:12:48 [block_pool.py:378] Successfully reset prefix cache

 19%|█▊        | 617/3300 [2:26:45<6:49:34,  9.16s/it]2026-04-12 13:12:57,207 | INFO | train_grpo_train | metrics_logged mode=train step=617

                                                      
{'loss': 0.0561, 'grad_norm': 11.658555030822754, 'learning_rate': 8.133333333333334e-06, 'num_tokens': 1241005.0, 'completions/mean_length': 50.5, 'completions/min_length': 45.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.5, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.4762023091316223, 'rewards/meter/std': 0.3319661617279053, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5425000190734863, 'rewards/judge_quality/std': 0.32805708050727844, 'rewards/repeat_penalty/mean': 0.9491168260574341, 'rewards/repeat_penalty/std': 0.04230201989412308, 'rewards/near_duplicate_penalty/mean': 0.966844916343689, 'rewards/near_duplicate_penalty/std': 0.035048648715019226, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9819197654724121, 'rewards/distinct_2/std': 0.03420789912343025, 'rewards/total_composite/mean': 0.27313536405563354, 'rewards/total_composite/std': 0.2842390835285187, 'reward': 0.27313536405563354, 'reward_std': 0.2842390835285187, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12931419909000397, 'sampling/sampling_logp_difference/max': 1.1296586990356445, 'sampling/importance_sampling_ratio/min': 0.32314351201057434, 'sampling/importance_sampling_ratio/mean': 1.0486185550689697, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1068106405436993, 'clip_ratio/low_mean': 0.07377674616873264, 'clip_ratio/low_min': 0.07377674616873264, 'clip_ratio/high_mean': 0.04960516653954983, 'clip_ratio/high_max': 0.04960516653954983, 'clip_ratio/region_mean': 0.12338191270828247, 'reward_total_mean': 0.27313536405563354, 'reward_meter_mean': 0.4762023091316223, 'reward_meter_std': 0.3319661617279053, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9491168260574341, 'reward_repeat_penalty_std': 0.04230201989412308, 'reward_near_duplicate_penalty_mean': 0.966844916343689, 'reward_near_duplicate_penalty_std': 0.035048648715019226, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9819197654724121, 'reward_distinct_2_std': 0.03420789912343025, 'reward_judge_quality_mean': 0.5425000190734863, 'reward_judge_quality_std': 0.32805708050727844, 'reward_total_composite_mean': 0.27313536405563354, 'reward_total_composite_std': 0.2842390835285187, 'epoch': 0.02}

 19%|█▊        | 617/3300 [2:26:45<6:49:34,  9.16s/it]INFO 04-12 13:12:57 [block_pool.py:378] Successfully reset prefix cache

 19%|█▊        | 618/3300 [2:27:00<8:11:19, 10.99s/it]2026-04-12 13:13:12,477 | INFO | train_grpo_train | metrics_logged mode=train step=618

                                                      
{'loss': -0.2818, 'grad_norm': 3.9043819904327393, 'learning_rate': 8.130303030303031e-06, 'num_tokens': 1243788.0, 'completions/mean_length': 208.875, 'completions/min_length': 54.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 165.57144165039062, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 221.0, 'rewards/meter/mean': 0.906912088394165, 'rewards/meter/std': 0.12399955838918686, 'rewards/count_adherence/mean': 0.734375, 'rewards/count_adherence/std': 0.3435470163822174, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9705565571784973, 'rewards/lexical_plausibility/std': 0.08327867090702057, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.14078859984874725, 'rewards/repeat_penalty/mean': 0.8192172646522522, 'rewards/repeat_penalty/std': 0.12516537308692932, 'rewards/near_duplicate_penalty/mean': 0.9616886377334595, 'rewards/near_duplicate_penalty/std': 0.023740138858556747, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.8889564275741577, 'rewards/distinct_2/std': 0.11105122417211533, 'rewards/total_composite/mean': 0.23532873392105103, 'rewards/total_composite/std': 0.1476844698190689, 'reward': 0.23532873392105103, 'reward_std': 0.14768445491790771, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15201258659362793, 'sampling/sampling_logp_difference/max': 2.5541534423828125, 'sampling/importance_sampling_ratio/min': 0.07775802910327911, 'sampling/importance_sampling_ratio/mean': 1.0186711549758911, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9685016870498657, 'clip_ratio/low_mean': 0.018518518656492233, 'clip_ratio/low_min': 0.018518518656492233, 'clip_ratio/high_mean': 0.110755305737257, 'clip_ratio/high_max': 0.110755305737257, 'clip_ratio/region_mean': 0.12927382439374924, 'reward_total_mean': 0.23532873392105103, 'reward_meter_mean': 0.906912088394165, 'reward_meter_std': 0.12399955838918686, 'reward_count_adherence_mean': 0.734375, 'reward_count_adherence_std': 0.3435470163822174, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9705565571784973, 'reward_lexical_plausibility_std': 0.08327867090702057, 'reward_repeat_penalty_mean': 0.8192172646522522, 'reward_repeat_penalty_std': 0.12516537308692932, 'reward_near_duplicate_penalty_mean': 0.9616886377334595, 'reward_near_duplicate_penalty_std': 0.023740138858556747, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.8889564275741577, 'reward_distinct_2_std': 0.11105122417211533, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.14078859984874725, 'reward_total_composite_mean': 0.23532873392105103, 'reward_total_composite_std': 0.1476844698190689, 'epoch': 0.02}

 19%|█▊        | 618/3300 [2:27:00<8:11:19, 10.99s/it]INFO 04-12 13:13:12 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 619/3300 [2:27:14<8:54:59, 11.97s/it]2026-04-12 13:13:26,739 | INFO | train_grpo_train | metrics_logged mode=train step=619

                                                      
{'loss': -0.0608, 'grad_norm': 2.523797035217285, 'learning_rate': 8.127272727272728e-06, 'num_tokens': 1245542.0, 'completions/mean_length': 297.25, 'completions/min_length': 64.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 82.5, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.3566693067550659, 'rewards/meter/std': 0.40195488929748535, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.18600596487522125, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/lexical_plausibility/mean': 0.8842828869819641, 'rewards/lexical_plausibility/std': 0.12774690985679626, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.21380899846553802, 'rewards/repeat_penalty/mean': 0.9717923402786255, 'rewards/repeat_penalty/std': 0.05927469581365585, 'rewards/near_duplicate_penalty/mean': 0.9900150299072266, 'rewards/near_duplicate_penalty/std': 0.008720648474991322, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9812382459640503, 'rewards/distinct_2/std': 0.05306617170572281, 'rewards/total_composite/mean': 0.08977677673101425, 'rewards/total_composite/std': 0.1531929075717926, 'reward': 0.08977677673101425, 'reward_std': 0.1531929075717926, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14239159226417542, 'sampling/sampling_logp_difference/max': 1.0537962913513184, 'sampling/importance_sampling_ratio/min': 0.3486118018627167, 'sampling/importance_sampling_ratio/mean': 1.039147973060608, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6592914760112762, 'clip_ratio/low_mean': 0.022849462926387787, 'clip_ratio/low_min': 0.022849462926387787, 'clip_ratio/high_mean': 0.0557304210960865, 'clip_ratio/high_max': 0.0557304210960865, 'clip_ratio/region_mean': 0.07857988402247429, 'reward_total_mean': 0.08977677673101425, 'reward_meter_mean': 0.3566693067550659, 'reward_meter_std': 0.40195488929748535, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.18600596487522125, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_lexical_plausibility_mean': 0.8842828869819641, 'reward_lexical_plausibility_std': 0.12774690985679626, 'reward_repeat_penalty_mean': 0.9717923402786255, 'reward_repeat_penalty_std': 0.05927469581365585, 'reward_near_duplicate_penalty_mean': 0.9900150299072266, 'reward_near_duplicate_penalty_std': 0.008720648474991322, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9812382459640503, 'reward_distinct_2_std': 0.05306617170572281, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.21380899846553802, 'reward_total_composite_mean': 0.08977677673101425, 'reward_total_composite_std': 0.1531929075717926, 'epoch': 0.02}

 19%|█▉        | 619/3300 [2:27:14<8:54:59, 11.97s/it]INFO 04-12 13:13:26 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 620/3300 [2:27:25<8:35:03, 11.53s/it]2026-04-12 13:13:37,236 | INFO | train_grpo_train | metrics_logged mode=train step=620

                                                      
{'loss': 0.0797, 'grad_norm': 7.215310096740723, 'learning_rate': 8.124242424242424e-06, 'num_tokens': 1248179.0, 'completions/mean_length': 131.625, 'completions/min_length': 112.0, 'completions/max_length': 149.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 131.625, 'completions/min_terminated_length': 112.0, 'completions/max_terminated_length': 149.0, 'rewards/meter/mean': 0.9546153545379639, 'rewards/meter/std': 0.05853186547756195, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.07715168595314026, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.21250000596046448, 'rewards/judge_quality/std': 0.09161253273487091, 'rewards/repeat_penalty/mean': 0.6516979932785034, 'rewards/repeat_penalty/std': 0.24182568490505219, 'rewards/near_duplicate_penalty/mean': 0.9193305969238281, 'rewards/near_duplicate_penalty/std': 0.07163485139608383, 'rewards/opening_diversity/mean': 0.987500011920929, 'rewards/opening_diversity/std': 0.0353553481400013, 'rewards/distinct_2/mean': 0.7071529030799866, 'rewards/distinct_2/std': 0.20859256386756897, 'rewards/total_composite/mean': 0.17269445955753326, 'rewards/total_composite/std': 0.061212316155433655, 'reward': 0.17269445955753326, 'reward_std': 0.061212316155433655, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14097511768341064, 'sampling/sampling_logp_difference/max': 2.991375207901001, 'sampling/importance_sampling_ratio/min': 0.05021832883358002, 'sampling/importance_sampling_ratio/mean': 1.019101858139038, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9945002570748329, 'clip_ratio/low_mean': 0.05802720983047038, 'clip_ratio/low_min': 0.05802720983047038, 'clip_ratio/high_mean': 0.049660032615065575, 'clip_ratio/high_max': 0.049660032615065575, 'clip_ratio/region_mean': 0.10768724244553596, 'reward_total_mean': 0.17269445955753326, 'reward_meter_mean': 0.9546153545379639, 'reward_meter_std': 0.05853186547756195, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.07715168595314026, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6516979932785034, 'reward_repeat_penalty_std': 0.24182568490505219, 'reward_near_duplicate_penalty_mean': 0.9193305969238281, 'reward_near_duplicate_penalty_std': 0.07163485139608383, 'reward_opening_diversity_mean': 0.987500011920929, 'reward_opening_diversity_std': 0.0353553481400013, 'reward_distinct_2_mean': 0.7071529030799866, 'reward_distinct_2_std': 0.20859256386756897, 'reward_judge_quality_mean': 0.21250000596046448, 'reward_judge_quality_std': 0.09161253273487091, 'reward_total_composite_mean': 0.17269445955753326, 'reward_total_composite_std': 0.061212316155433655, 'epoch': 0.02}

 19%|█▉        | 620/3300 [2:27:25<8:35:03, 11.53s/it]INFO 04-12 13:13:37 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 621/3300 [2:27:38<9:01:54, 12.14s/it]2026-04-12 13:13:50,787 | INFO | train_grpo_train | metrics_logged mode=train step=621

                                                      
{'loss': -0.0109, 'grad_norm': 6.699946880340576, 'learning_rate': 8.121212121212121e-06, 'num_tokens': 1249647.0, 'completions/mean_length': 98.5, 'completions/min_length': 36.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 39.42857360839844, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.16228392720222473, 'rewards/meter/std': 0.1424250602722168, '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/lexical_plausibility/mean': 0.9783446192741394, 'rewards/lexical_plausibility/std': 0.06125066801905632, 'rewards/judge_quality/mean': 0.6725000143051147, 'rewards/judge_quality/std': 0.32459867000579834, 'rewards/repeat_penalty/mean': 0.9503357410430908, 'rewards/repeat_penalty/std': 0.08662397414445877, 'rewards/near_duplicate_penalty/mean': 0.991681694984436, 'rewards/near_duplicate_penalty/std': 0.018021628260612488, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9898785352706909, 'rewards/distinct_2/std': 0.028627805411815643, 'rewards/total_composite/mean': 0.13230977952480316, 'rewards/total_composite/std': 0.1309736967086792, 'reward': 0.13230977952480316, 'reward_std': 0.1309736967086792, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16338719427585602, 'sampling/sampling_logp_difference/max': 4.11427116394043, 'sampling/importance_sampling_ratio/min': 0.016337843611836433, 'sampling/importance_sampling_ratio/mean': 1.0064324140548706, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6377340257167816, 'clip_ratio/low_mean': 0.061812132596969604, 'clip_ratio/low_min': 0.061812132596969604, 'clip_ratio/high_mean': 0.03415915835648775, 'clip_ratio/high_max': 0.03415915835648775, 'clip_ratio/region_mean': 0.09597129095345736, 'reward_total_mean': 0.13230977952480316, 'reward_meter_mean': 0.16228392720222473, 'reward_meter_std': 0.1424250602722168, '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_lexical_plausibility_mean': 0.9783446192741394, 'reward_lexical_plausibility_std': 0.06125066801905632, 'reward_repeat_penalty_mean': 0.9503357410430908, 'reward_repeat_penalty_std': 0.08662397414445877, 'reward_near_duplicate_penalty_mean': 0.991681694984436, 'reward_near_duplicate_penalty_std': 0.018021628260612488, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9898785352706909, 'reward_distinct_2_std': 0.028627805411815643, 'reward_judge_quality_mean': 0.6725000143051147, 'reward_judge_quality_std': 0.32459867000579834, 'reward_total_composite_mean': 0.13230977952480316, 'reward_total_composite_std': 0.1309736967086792, 'epoch': 0.02}

 19%|█▉        | 621/3300 [2:27:38<9:01:54, 12.14s/it]INFO 04-12 13:13:51 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 622/3300 [2:27:53<9:33:10, 12.84s/it]2026-04-12 13:14:05,278 | INFO | train_grpo_train | metrics_logged mode=train step=622

                                                      
{'loss': -0.0424, 'grad_norm': 2.576357841491699, 'learning_rate': 8.118181818181819e-06, 'num_tokens': 1251084.0, 'completions/mean_length': 159.625, 'completions/min_length': 37.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 42.16666793823242, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.6240122318267822, 'rewards/meter/std': 0.41501474380493164, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9315087795257568, 'rewards/lexical_plausibility/std': 0.10194209218025208, 'rewards/judge_quality/mean': 0.45499998331069946, 'rewards/judge_quality/std': 0.33200690150260925, 'rewards/repeat_penalty/mean': 0.741641640663147, 'rewards/repeat_penalty/std': 0.2957896590232849, 'rewards/near_duplicate_penalty/mean': 0.9325931668281555, 'rewards/near_duplicate_penalty/std': 0.07531813532114029, 'rewards/opening_diversity/mean': 0.828125, 'rewards/opening_diversity/std': 0.22097088396549225, 'rewards/distinct_2/mean': 0.9144536256790161, 'rewards/distinct_2/std': 0.120970219373703, 'rewards/total_composite/mean': 0.32535994052886963, 'rewards/total_composite/std': 0.3713827431201935, 'reward': 0.32535994052886963, 'reward_std': 0.3713827431201935, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16923077404499054, 'sampling/sampling_logp_difference/max': 1.417853832244873, 'sampling/importance_sampling_ratio/min': 0.24223333597183228, 'sampling/importance_sampling_ratio/mean': 1.0087859630584717, 'sampling/importance_sampling_ratio/max': 1.892889380455017, 'entropy': 1.0709325671195984, 'clip_ratio/low_mean': 0.0752748791128397, 'clip_ratio/low_min': 0.0752748791128397, 'clip_ratio/high_mean': 0.032917533069849014, 'clip_ratio/high_max': 0.032917533069849014, 'clip_ratio/region_mean': 0.10819241218268871, 'reward_total_mean': 0.32535994052886963, 'reward_meter_mean': 0.6240122318267822, 'reward_meter_std': 0.41501474380493164, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9315087795257568, 'reward_lexical_plausibility_std': 0.10194209218025208, 'reward_repeat_penalty_mean': 0.741641640663147, 'reward_repeat_penalty_std': 0.2957896590232849, 'reward_near_duplicate_penalty_mean': 0.9325931668281555, 'reward_near_duplicate_penalty_std': 0.07531813532114029, 'reward_opening_diversity_mean': 0.828125, 'reward_opening_diversity_std': 0.22097088396549225, 'reward_distinct_2_mean': 0.9144536256790161, 'reward_distinct_2_std': 0.120970219373703, 'reward_judge_quality_mean': 0.45499998331069946, 'reward_judge_quality_std': 0.33200690150260925, 'reward_total_composite_mean': 0.32535994052886963, 'reward_total_composite_std': 0.3713827431201935, 'epoch': 0.02}

 19%|█▉        | 622/3300 [2:27:53<9:33:10, 12.84s/it]INFO 04-12 13:14:05 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 623/3300 [2:28:07<9:47:06, 13.16s/it]2026-04-12 13:14:19,177 | INFO | train_grpo_train | metrics_logged mode=train step=623

                                                      
{'loss': -0.0588, 'grad_norm': 1.6638461351394653, 'learning_rate': 8.115151515151516e-06, 'num_tokens': 1252571.0, 'completions/mean_length': 276.875, 'completions/min_length': 32.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 41.75, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.5826632380485535, 'rewards/meter/std': 0.355167418718338, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/lexical_plausibility/mean': 0.869391679763794, 'rewards/lexical_plausibility/std': 0.15019075572490692, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.19086270034313202, 'rewards/repeat_penalty/mean': 0.9958506226539612, 'rewards/repeat_penalty/std': 0.008159920573234558, 'rewards/near_duplicate_penalty/mean': 0.9958506226539612, 'rewards/near_duplicate_penalty/std': 0.008159920573234558, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.13165487349033356, 'rewards/total_composite/std': 0.17305083572864532, 'reward': 0.13165487349033356, 'reward_std': 0.17305083572864532, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17046697437763214, 'sampling/sampling_logp_difference/max': 1.659440040588379, 'sampling/importance_sampling_ratio/min': 0.19024547934532166, 'sampling/importance_sampling_ratio/mean': 1.0163774490356445, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6407098323106766, 'clip_ratio/low_mean': 0.016025641933083534, 'clip_ratio/low_min': 0.016025641933083534, 'clip_ratio/high_mean': 0.05231627356261015, 'clip_ratio/high_max': 0.05231627356261015, 'clip_ratio/region_mean': 0.06834191549569368, 'reward_total_mean': 0.13165487349033356, 'reward_meter_mean': 0.5826632380485535, 'reward_meter_std': 0.355167418718338, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_lexical_plausibility_mean': 0.869391679763794, 'reward_lexical_plausibility_std': 0.15019075572490692, 'reward_repeat_penalty_mean': 0.9958506226539612, 'reward_repeat_penalty_std': 0.008159920573234558, 'reward_near_duplicate_penalty_mean': 0.9958506226539612, 'reward_near_duplicate_penalty_std': 0.008159920573234558, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.19086270034313202, 'reward_total_composite_mean': 0.13165487349033356, 'reward_total_composite_std': 0.17305083572864532, 'epoch': 0.03}

 19%|█▉        | 623/3300 [2:28:07<9:47:06, 13.16s/it]INFO 04-12 13:14:19 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 624/3300 [2:28:20<9:49:36, 13.22s/it]2026-04-12 13:14:32,534 | INFO | train_grpo_train | metrics_logged mode=train step=624

                                                      
{'loss': -0.0939, 'grad_norm': 2.0242741107940674, 'learning_rate': 8.112121212121213e-06, 'num_tokens': 1254095.0, 'completions/mean_length': 169.5, 'completions/min_length': 49.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 55.333335876464844, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.920796275138855, 'rewards/meter/std': 0.1217091754078865, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9258214235305786, 'rewards/lexical_plausibility/std': 0.12028312683105469, 'rewards/judge_quality/mean': 0.3787499964237213, 'rewards/judge_quality/std': 0.2630283832550049, 'rewards/repeat_penalty/mean': 0.7558110952377319, 'rewards/repeat_penalty/std': 0.2434026151895523, 'rewards/near_duplicate_penalty/mean': 0.9661861658096313, 'rewards/near_duplicate_penalty/std': 0.04693194851279259, 'rewards/opening_diversity/mean': 0.84375, 'rewards/opening_diversity/std': 0.18600596487522125, 'rewards/distinct_2/mean': 0.9123941659927368, 'rewards/distinct_2/std': 0.10995517671108246, 'rewards/total_composite/mean': 0.3570767641067505, 'rewards/total_composite/std': 0.27805736660957336, 'reward': 0.3570767641067505, 'reward_std': 0.27805736660957336, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16967812180519104, 'sampling/sampling_logp_difference/max': 1.991382122039795, 'sampling/importance_sampling_ratio/min': 0.13650661706924438, 'sampling/importance_sampling_ratio/mean': 1.0271700620651245, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.095145545899868, 'clip_ratio/low_mean': 0.02461662609130144, 'clip_ratio/low_min': 0.02461662609130144, 'clip_ratio/high_mean': 0.06225738022476435, 'clip_ratio/high_max': 0.06225738022476435, 'clip_ratio/region_mean': 0.08687400631606579, 'reward_total_mean': 0.3570767641067505, 'reward_meter_mean': 0.920796275138855, 'reward_meter_std': 0.1217091754078865, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9258214235305786, 'reward_lexical_plausibility_std': 0.12028312683105469, 'reward_repeat_penalty_mean': 0.7558110952377319, 'reward_repeat_penalty_std': 0.2434026151895523, 'reward_near_duplicate_penalty_mean': 0.9661861658096313, 'reward_near_duplicate_penalty_std': 0.04693194851279259, 'reward_opening_diversity_mean': 0.84375, 'reward_opening_diversity_std': 0.18600596487522125, 'reward_distinct_2_mean': 0.9123941659927368, 'reward_distinct_2_std': 0.10995517671108246, 'reward_judge_quality_mean': 0.3787499964237213, 'reward_judge_quality_std': 0.2630283832550049, 'reward_total_composite_mean': 0.3570767641067505, 'reward_total_composite_std': 0.27805736660957336, 'epoch': 0.03}

 19%|█▉        | 624/3300 [2:28:20<9:49:36, 13.22s/it]INFO 04-12 13:14:32 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 625/3300 [2:28:34<10:00:19, 13.47s/it]2026-04-12 13:14:46,574 | INFO | train_grpo_train | metrics_logged mode=train step=625

                                                       
{'loss': -0.0406, 'grad_norm': 4.97029447555542, 'learning_rate': 8.10909090909091e-06, 'num_tokens': 1255726.0, 'completions/mean_length': 107.875, 'completions/min_length': 46.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 50.142860412597656, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.5163154602050781, 'rewards/meter/std': 0.3986039161682129, '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/lexical_plausibility/mean': 0.9473627805709839, 'rewards/lexical_plausibility/std': 0.11657222360372543, 'rewards/judge_quality/mean': 0.34999996423721313, 'rewards/judge_quality/std': 0.18516401946544647, 'rewards/repeat_penalty/mean': 0.9596666097640991, 'rewards/repeat_penalty/std': 0.036881618201732635, 'rewards/near_duplicate_penalty/mean': 0.9753938913345337, 'rewards/near_duplicate_penalty/std': 0.02577846311032772, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9839464426040649, 'rewards/distinct_2/std': 0.029759714379906654, 'rewards/total_composite/mean': 0.23325850069522858, 'rewards/total_composite/std': 0.22179822623729706, 'reward': 0.23325850069522858, 'reward_std': 0.22179822623729706, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18569746613502502, 'sampling/sampling_logp_difference/max': 1.4033632278442383, 'sampling/importance_sampling_ratio/min': 0.24576899409294128, 'sampling/importance_sampling_ratio/mean': 1.0396032333374023, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9483740329742432, 'clip_ratio/low_mean': 0.04591658432036638, 'clip_ratio/low_min': 0.04591658432036638, 'clip_ratio/high_mean': 0.09515212289988995, 'clip_ratio/high_max': 0.09515212289988995, 'clip_ratio/region_mean': 0.14106870722025633, 'reward_total_mean': 0.23325850069522858, 'reward_meter_mean': 0.5163154602050781, 'reward_meter_std': 0.3986039161682129, '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_lexical_plausibility_mean': 0.9473627805709839, 'reward_lexical_plausibility_std': 0.11657222360372543, 'reward_repeat_penalty_mean': 0.9596666097640991, 'reward_repeat_penalty_std': 0.036881618201732635, 'reward_near_duplicate_penalty_mean': 0.9753938913345337, 'reward_near_duplicate_penalty_std': 0.02577846311032772, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9839464426040649, 'reward_distinct_2_std': 0.029759714379906654, 'reward_judge_quality_mean': 0.34999996423721313, 'reward_judge_quality_std': 0.18516401946544647, 'reward_total_composite_mean': 0.23325850069522858, 'reward_total_composite_std': 0.22179822623729706, 'epoch': 0.03}

 19%|█▉        | 625/3300 [2:28:34<10:00:19, 13.47s/it]INFO 04-12 13:14:46 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 626/3300 [2:28:48<10:11:27, 13.72s/it]2026-04-12 13:15:00,889 | INFO | train_grpo_train | metrics_logged mode=train step=626

                                                       
{'loss': -0.0912, 'grad_norm': 2.7027525901794434, 'learning_rate': 8.106060606060606e-06, 'num_tokens': 1257290.0, 'completions/mean_length': 231.5, 'completions/min_length': 59.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 63.20000076293945, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.6179291605949402, 'rewards/meter/std': 0.42140281200408936, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.2519763112068176, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.9073654413223267, 'rewards/lexical_plausibility/std': 0.13420233130455017, 'rewards/judge_quality/mean': 0.3787499964237213, 'rewards/judge_quality/std': 0.32166698575019836, 'rewards/repeat_penalty/mean': 0.9481266736984253, 'rewards/repeat_penalty/std': 0.051743339747190475, 'rewards/near_duplicate_penalty/mean': 0.9822664260864258, 'rewards/near_duplicate_penalty/std': 0.021032456308603287, 'rewards/opening_diversity/mean': 0.987500011920929, 'rewards/opening_diversity/std': 0.0353553481400013, 'rewards/distinct_2/mean': 0.9772042036056519, 'rewards/distinct_2/std': 0.02437390573322773, 'rewards/total_composite/mean': 0.30892783403396606, 'rewards/total_composite/std': 0.34324121475219727, 'reward': 0.30892783403396606, 'reward_std': 0.34324121475219727, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15500257909297943, 'sampling/sampling_logp_difference/max': 1.173779010772705, 'sampling/importance_sampling_ratio/min': 0.3091962933540344, 'sampling/importance_sampling_ratio/mean': 1.0177265405654907, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9733167588710785, 'clip_ratio/low_mean': 0.029982023872435093, 'clip_ratio/low_min': 0.029982023872435093, 'clip_ratio/high_mean': 0.06459018401801586, 'clip_ratio/high_max': 0.06459018401801586, 'clip_ratio/region_mean': 0.09457220789045095, 'reward_total_mean': 0.30892783403396606, 'reward_meter_mean': 0.6179291605949402, 'reward_meter_std': 0.42140281200408936, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.2519763112068176, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.9073654413223267, 'reward_lexical_plausibility_std': 0.13420233130455017, 'reward_repeat_penalty_mean': 0.9481266736984253, 'reward_repeat_penalty_std': 0.051743339747190475, 'reward_near_duplicate_penalty_mean': 0.9822664260864258, 'reward_near_duplicate_penalty_std': 0.021032456308603287, 'reward_opening_diversity_mean': 0.987500011920929, 'reward_opening_diversity_std': 0.0353553481400013, 'reward_distinct_2_mean': 0.9772042036056519, 'reward_distinct_2_std': 0.02437390573322773, 'reward_judge_quality_mean': 0.3787499964237213, 'reward_judge_quality_std': 0.32166698575019836, 'reward_total_composite_mean': 0.30892783403396606, 'reward_total_composite_std': 0.34324121475219727, 'epoch': 0.03}

 19%|█▉        | 626/3300 [2:28:48<10:11:27, 13.72s/it]INFO 04-12 13:15:01 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 627/3300 [2:28:58<9:13:23, 12.42s/it] 2026-04-12 13:15:10,286 | INFO | train_grpo_train | metrics_logged mode=train step=627

                                                      
{'loss': -0.0331, 'grad_norm': 13.08083438873291, 'learning_rate': 8.103030303030303e-06, 'num_tokens': 1258791.0, 'completions/mean_length': 44.625, 'completions/min_length': 38.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.625, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.9668298959732056, 'rewards/meter/std': 0.03126410022377968, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4675000309944153, 'rewards/judge_quality/std': 0.29261142015457153, 'rewards/repeat_penalty/mean': 0.9204707145690918, 'rewards/repeat_penalty/std': 0.12259316444396973, 'rewards/near_duplicate_penalty/mean': 0.9745610356330872, 'rewards/near_duplicate_penalty/std': 0.021819118410348892, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9720279574394226, 'rewards/distinct_2/std': 0.05644134432077408, 'rewards/total_composite/mean': 0.3993763029575348, 'rewards/total_composite/std': 0.325158953666687, 'reward': 0.3993763029575348, 'reward_std': 0.3251589238643646, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12139907479286194, 'sampling/sampling_logp_difference/max': 1.9642690420150757, 'sampling/importance_sampling_ratio/min': 0.14025837182998657, 'sampling/importance_sampling_ratio/mean': 1.0100677013397217, 'sampling/importance_sampling_ratio/max': 1.6912506818771362, 'entropy': 0.8464110419154167, 'clip_ratio/low_mean': 0.0733287027105689, 'clip_ratio/low_min': 0.0733287027105689, 'clip_ratio/high_mean': 0.044492464512586594, 'clip_ratio/high_max': 0.044492464512586594, 'clip_ratio/region_mean': 0.1178211672231555, 'reward_total_mean': 0.3993763029575348, 'reward_meter_mean': 0.9668298959732056, 'reward_meter_std': 0.03126410022377968, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9204707145690918, 'reward_repeat_penalty_std': 0.12259316444396973, 'reward_near_duplicate_penalty_mean': 0.9745610356330872, 'reward_near_duplicate_penalty_std': 0.021819118410348892, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9720279574394226, 'reward_distinct_2_std': 0.05644134432077408, 'reward_judge_quality_mean': 0.4675000309944153, 'reward_judge_quality_std': 0.29261142015457153, 'reward_total_composite_mean': 0.3993763029575348, 'reward_total_composite_std': 0.325158953666687, 'epoch': 0.03}

 19%|█▉        | 627/3300 [2:28:58<9:13:23, 12.42s/it]INFO 04-12 13:15:10 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 628/3300 [2:29:12<9:42:03, 13.07s/it]2026-04-12 13:15:24,863 | INFO | train_grpo_train | metrics_logged mode=train step=628

                                                      
{'loss': -0.0986, 'grad_norm': 2.82222843170166, 'learning_rate': 8.1e-06, 'num_tokens': 1260449.0, 'completions/mean_length': 175.25, 'completions/min_length': 54.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 63.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.6874505281448364, 'rewards/meter/std': 0.4127632975578308, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.9381188750267029, 'rewards/lexical_plausibility/std': 0.11458229273557663, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.18516401946544647, 'rewards/repeat_penalty/mean': 0.7721683382987976, 'rewards/repeat_penalty/std': 0.23168954253196716, 'rewards/near_duplicate_penalty/mean': 0.9404508471488953, 'rewards/near_duplicate_penalty/std': 0.05880289152264595, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.88327556848526, 'rewards/distinct_2/std': 0.12183911353349686, 'rewards/total_composite/mean': 0.23909616470336914, 'rewards/total_composite/std': 0.2025579810142517, 'reward': 0.23909616470336914, 'reward_std': 0.2025579810142517, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14117272198200226, 'sampling/sampling_logp_difference/max': 1.0313647985458374, 'sampling/importance_sampling_ratio/min': 0.3565200865268707, 'sampling/importance_sampling_ratio/mean': 1.0376149415969849, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8580816984176636, 'clip_ratio/low_mean': 0.02217741869390011, 'clip_ratio/low_min': 0.02217741869390011, 'clip_ratio/high_mean': 0.07609720341861248, 'clip_ratio/high_max': 0.07609720341861248, 'clip_ratio/region_mean': 0.09827462211251259, 'reward_total_mean': 0.23909616470336914, 'reward_meter_mean': 0.6874505281448364, 'reward_meter_std': 0.4127632975578308, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.9381188750267029, 'reward_lexical_plausibility_std': 0.11458229273557663, 'reward_repeat_penalty_mean': 0.7721683382987976, 'reward_repeat_penalty_std': 0.23168954253196716, 'reward_near_duplicate_penalty_mean': 0.9404508471488953, 'reward_near_duplicate_penalty_std': 0.05880289152264595, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.88327556848526, 'reward_distinct_2_std': 0.12183911353349686, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.18516401946544647, 'reward_total_composite_mean': 0.23909616470336914, 'reward_total_composite_std': 0.2025579810142517, 'epoch': 0.03}

 19%|█▉        | 628/3300 [2:29:12<9:42:03, 13.07s/it]INFO 04-12 13:15:25 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 629/3300 [2:29:26<9:51:36, 13.29s/it]2026-04-12 13:15:38,668 | INFO | train_grpo_train | metrics_logged mode=train step=629

                                                      
{'loss': -0.0524, 'grad_norm': 1.4695098400115967, 'learning_rate': 8.096969696969698e-06, 'num_tokens': 1261783.0, 'completions/mean_length': 336.75, 'completions/min_length': 37.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.625, 'completions/mean_terminated_length': 44.66666793823242, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.24681605398654938, 'rewards/meter/std': 0.3352571427822113, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.375, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.8966175317764282, 'rewards/lexical_plausibility/std': 0.08834678679704666, 'rewards/judge_quality/mean': 0.20000000298023224, 'rewards/judge_quality/std': 0.26592159271240234, 'rewards/repeat_penalty/mean': 0.9511940479278564, 'rewards/repeat_penalty/std': 0.08496034890413284, 'rewards/near_duplicate_penalty/mean': 0.9906213879585266, 'rewards/near_duplicate_penalty/std': 0.01871313527226448, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9918146133422852, 'rewards/distinct_2/std': 0.02160203829407692, 'rewards/total_composite/mean': 0.03236179053783417, 'rewards/total_composite/std': 0.04901016503572464, 'reward': 0.03236179053783417, 'reward_std': 0.04901016876101494, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23172251880168915, 'sampling/sampling_logp_difference/max': 1.4877829551696777, 'sampling/importance_sampling_ratio/min': 0.22587287425994873, 'sampling/importance_sampling_ratio/mean': 1.064292073249817, 'sampling/importance_sampling_ratio/max': 1.9341951608657837, 'entropy': 0.8464368432760239, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0490370187908411, 'clip_ratio/high_max': 0.0490370187908411, 'clip_ratio/region_mean': 0.0490370187908411, 'reward_total_mean': 0.03236179053783417, 'reward_meter_mean': 0.24681605398654938, 'reward_meter_std': 0.3352571427822113, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.375, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.8966175317764282, 'reward_lexical_plausibility_std': 0.08834678679704666, 'reward_repeat_penalty_mean': 0.9511940479278564, 'reward_repeat_penalty_std': 0.08496034890413284, 'reward_near_duplicate_penalty_mean': 0.9906213879585266, 'reward_near_duplicate_penalty_std': 0.01871313527226448, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9918146133422852, 'reward_distinct_2_std': 0.02160203829407692, 'reward_judge_quality_mean': 0.20000000298023224, 'reward_judge_quality_std': 0.26592159271240234, 'reward_total_composite_mean': 0.03236179053783417, 'reward_total_composite_std': 0.04901016503572464, 'epoch': 0.03}

 19%|█▉        | 629/3300 [2:29:26<9:51:36, 13.29s/it]INFO 04-12 13:15:38 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 630/3300 [2:29:42<10:20:39, 13.95s/it]2026-04-12 13:15:54,148 | INFO | train_grpo_train | metrics_logged mode=train step=630

                                                       
{'loss': -0.0517, 'grad_norm': 5.594262599945068, 'learning_rate': 8.093939393939395e-06, 'num_tokens': 1263672.0, 'completions/mean_length': 122.125, 'completions/min_length': 61.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 66.42857360839844, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.5022074580192566, 'rewards/meter/std': 0.3894704282283783, '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/lexical_plausibility/mean': 0.9602227210998535, 'rewards/lexical_plausibility/std': 0.11250722408294678, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.14880475401878357, 'rewards/repeat_penalty/mean': 0.8946222066879272, 'rewards/repeat_penalty/std': 0.0831746757030487, 'rewards/near_duplicate_penalty/mean': 0.9494065642356873, 'rewards/near_duplicate_penalty/std': 0.03705008327960968, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9406620264053345, 'rewards/distinct_2/std': 0.05450836196541786, 'rewards/total_composite/mean': 0.1588403284549713, 'rewards/total_composite/std': 0.15504039824008942, 'reward': 0.1588403284549713, 'reward_std': 0.15504039824008942, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20285888016223907, 'sampling/sampling_logp_difference/max': 1.7668914794921875, 'sampling/importance_sampling_ratio/min': 0.1708633005619049, 'sampling/importance_sampling_ratio/mean': 1.0529505014419556, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.576981171965599, 'clip_ratio/low_mean': 0.059718722477555275, 'clip_ratio/low_min': 0.059718722477555275, 'clip_ratio/high_mean': 0.0863551665097475, 'clip_ratio/high_max': 0.0863551665097475, 'clip_ratio/region_mean': 0.14607388898730278, 'reward_total_mean': 0.1588403284549713, 'reward_meter_mean': 0.5022074580192566, 'reward_meter_std': 0.3894704282283783, '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_lexical_plausibility_mean': 0.9602227210998535, 'reward_lexical_plausibility_std': 0.11250722408294678, 'reward_repeat_penalty_mean': 0.8946222066879272, 'reward_repeat_penalty_std': 0.0831746757030487, 'reward_near_duplicate_penalty_mean': 0.9494065642356873, 'reward_near_duplicate_penalty_std': 0.03705008327960968, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9406620264053345, 'reward_distinct_2_std': 0.05450836196541786, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.14880475401878357, 'reward_total_composite_mean': 0.1588403284549713, 'reward_total_composite_std': 0.15504039824008942, 'epoch': 0.03}

 19%|█▉        | 630/3300 [2:29:42<10:20:39, 13.95s/it]INFO 04-12 13:15:54 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 631/3300 [2:29:56<10:20:49, 13.96s/it]2026-04-12 13:16:08,126 | INFO | train_grpo_train | metrics_logged mode=train step=631

                                                       
{'loss': -0.0663, 'grad_norm': 5.523833274841309, 'learning_rate': 8.090909090909092e-06, 'num_tokens': 1265834.0, 'completions/mean_length': 162.25, 'completions/min_length': 104.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 112.28572082519531, 'completions/min_terminated_length': 104.0, 'completions/max_terminated_length': 127.0, 'rewards/meter/mean': 0.9624959826469421, 'rewards/meter/std': 0.035045068711042404, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9842867851257324, 'rewards/lexical_plausibility/std': 0.04444364085793495, 'rewards/judge_quality/mean': 0.32749998569488525, 'rewards/judge_quality/std': 0.27343058586120605, 'rewards/repeat_penalty/mean': 0.8556100130081177, 'rewards/repeat_penalty/std': 0.1390679031610489, 'rewards/near_duplicate_penalty/mean': 0.9416369199752808, 'rewards/near_duplicate_penalty/std': 0.035905905067920685, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9056022763252258, 'rewards/distinct_2/std': 0.12288402020931244, 'rewards/total_composite/mean': 0.30060014128685, 'rewards/total_composite/std': 0.28588518500328064, 'reward': 0.30060014128685, 'reward_std': 0.28588518500328064, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1634335070848465, 'sampling/sampling_logp_difference/max': 1.9987869262695312, 'sampling/importance_sampling_ratio/min': 0.13549955189228058, 'sampling/importance_sampling_ratio/mean': 1.0251843929290771, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1761610209941864, 'clip_ratio/low_mean': 0.03800287377089262, 'clip_ratio/low_min': 0.03800287377089262, 'clip_ratio/high_mean': 0.10151827149093151, 'clip_ratio/high_max': 0.10151827149093151, 'clip_ratio/region_mean': 0.13952114526182413, 'reward_total_mean': 0.30060014128685, 'reward_meter_mean': 0.9624959826469421, 'reward_meter_std': 0.035045068711042404, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9842867851257324, 'reward_lexical_plausibility_std': 0.04444364085793495, 'reward_repeat_penalty_mean': 0.8556100130081177, 'reward_repeat_penalty_std': 0.1390679031610489, 'reward_near_duplicate_penalty_mean': 0.9416369199752808, 'reward_near_duplicate_penalty_std': 0.035905905067920685, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9056022763252258, 'reward_distinct_2_std': 0.12288402020931244, 'reward_judge_quality_mean': 0.32749998569488525, 'reward_judge_quality_std': 0.27343058586120605, 'reward_total_composite_mean': 0.30060014128685, 'reward_total_composite_std': 0.28588518500328064, 'epoch': 0.03}

 19%|█▉        | 631/3300 [2:29:56<10:20:49, 13.96s/it]INFO 04-12 13:16:08 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 632/3300 [2:30:09<10:18:14, 13.90s/it]2026-04-12 13:16:21,906 | INFO | train_grpo_train | metrics_logged mode=train step=632

                                                       
{'loss': -0.0608, 'grad_norm': 2.5967648029327393, 'learning_rate': 8.08787878787879e-06, 'num_tokens': 1267300.0, 'completions/mean_length': 215.25, 'completions/min_length': 32.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 37.20000076293945, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.261963814496994, 'rewards/meter/std': 0.3052437901496887, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9159097075462341, 'rewards/lexical_plausibility/std': 0.11721618473529816, 'rewards/judge_quality/mean': 0.23749998211860657, 'rewards/judge_quality/std': 0.18077215552330017, 'rewards/repeat_penalty/mean': 0.9788684844970703, 'rewards/repeat_penalty/std': 0.0204145647585392, 'rewards/near_duplicate_penalty/mean': 0.9788684844970703, 'rewards/near_duplicate_penalty/std': 0.0204145647585392, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.04824015125632286, 'rewards/total_composite/std': 0.04420493170619011, 'reward': 0.04824015125632286, 'reward_std': 0.04420492798089981, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21038556098937988, 'sampling/sampling_logp_difference/max': 1.7036457061767578, 'sampling/importance_sampling_ratio/min': 0.1820187270641327, 'sampling/importance_sampling_ratio/mean': 1.0416951179504395, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0705594271421432, 'clip_ratio/low_mean': 0.01171875, 'clip_ratio/low_min': 0.01171875, 'clip_ratio/high_mean': 0.09922089613974094, 'clip_ratio/high_max': 0.09922089613974094, 'clip_ratio/region_mean': 0.11093964613974094, 'reward_total_mean': 0.04824015125632286, 'reward_meter_mean': 0.261963814496994, 'reward_meter_std': 0.3052437901496887, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9159097075462341, 'reward_lexical_plausibility_std': 0.11721618473529816, 'reward_repeat_penalty_mean': 0.9788684844970703, 'reward_repeat_penalty_std': 0.0204145647585392, 'reward_near_duplicate_penalty_mean': 0.9788684844970703, 'reward_near_duplicate_penalty_std': 0.0204145647585392, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.23749998211860657, 'reward_judge_quality_std': 0.18077215552330017, 'reward_total_composite_mean': 0.04824015125632286, 'reward_total_composite_std': 0.04420493170619011, 'epoch': 0.03}

 19%|█▉        | 632/3300 [2:30:09<10:18:14, 13.90s/it]INFO 04-12 13:16:22 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 633/3300 [2:30:17<8:56:31, 12.07s/it] 2026-04-12 13:16:29,700 | INFO | train_grpo_train | metrics_logged mode=train step=633

                                                      
{'loss': -0.0191, 'grad_norm': 14.925973892211914, 'learning_rate': 8.084848484848485e-06, 'num_tokens': 1268979.0, 'completions/mean_length': 43.875, 'completions/min_length': 39.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.875, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.5629684925079346, 'rewards/meter/std': 0.4164302349090576, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7762500047683716, 'rewards/judge_quality/std': 0.30500292778015137, 'rewards/repeat_penalty/mean': 0.9292605519294739, 'rewards/repeat_penalty/std': 0.1500680148601532, 'rewards/near_duplicate_penalty/mean': 0.9707680940628052, 'rewards/near_duplicate_penalty/std': 0.03805668279528618, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9797570705413818, 'rewards/distinct_2/std': 0.05725561082363129, 'rewards/total_composite/mean': 0.39023247361183167, 'rewards/total_composite/std': 0.3673883080482483, 'reward': 0.39023247361183167, 'reward_std': 0.3673883378505707, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14959272742271423, 'sampling/sampling_logp_difference/max': 2.071025848388672, 'sampling/importance_sampling_ratio/min': 0.12605640292167664, 'sampling/importance_sampling_ratio/mean': 1.0124744176864624, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2291499078273773, 'clip_ratio/low_mean': 0.05689174123108387, 'clip_ratio/low_min': 0.05689174123108387, 'clip_ratio/high_mean': 0.06366459582932293, 'clip_ratio/high_max': 0.06366459582932293, 'clip_ratio/region_mean': 0.1205563370604068, 'reward_total_mean': 0.39023247361183167, 'reward_meter_mean': 0.5629684925079346, 'reward_meter_std': 0.4164302349090576, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9292605519294739, 'reward_repeat_penalty_std': 0.1500680148601532, 'reward_near_duplicate_penalty_mean': 0.9707680940628052, 'reward_near_duplicate_penalty_std': 0.03805668279528618, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9797570705413818, 'reward_distinct_2_std': 0.05725561082363129, 'reward_judge_quality_mean': 0.7762500047683716, 'reward_judge_quality_std': 0.30500292778015137, 'reward_total_composite_mean': 0.39023247361183167, 'reward_total_composite_std': 0.3673883080482483, 'epoch': 0.03}

 19%|█▉        | 633/3300 [2:30:17<8:56:31, 12.07s/it]INFO 04-12 13:16:29 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 634/3300 [2:30:31<9:21:33, 12.64s/it]2026-04-12 13:16:43,661 | INFO | train_grpo_train | metrics_logged mode=train step=634

                                                      
{'loss': -0.0168, 'grad_norm': 1.4095944166183472, 'learning_rate': 8.081818181818182e-06, 'num_tokens': 1270272.0, 'completions/mean_length': 265.625, 'completions/min_length': 15.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 19.25, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.2789691686630249, 'rewards/meter/std': 0.28043612837791443, '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/lexical_plausibility/mean': 0.8315871357917786, 'rewards/lexical_plausibility/std': 0.1664452999830246, 'rewards/judge_quality/mean': 0.2862499952316284, 'rewards/judge_quality/std': 0.3945680260658264, 'rewards/repeat_penalty/mean': 0.7809751033782959, 'rewards/repeat_penalty/std': 0.08850277960300446, 'rewards/near_duplicate_penalty/mean': 0.9996334314346313, 'rewards/near_duplicate_penalty/std': 0.0010368125513195992, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1212349608540535, 'rewards/total_composite/std': 0.2543593645095825, 'reward': 0.1212349608540535, 'reward_std': 0.2543593645095825, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17294777929782867, 'sampling/sampling_logp_difference/max': 1.9102692604064941, 'sampling/importance_sampling_ratio/min': 0.1480405181646347, 'sampling/importance_sampling_ratio/mean': 1.0439677238464355, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7391538321971893, 'clip_ratio/low_mean': 0.03611111221835017, 'clip_ratio/low_min': 0.03611111221835017, 'clip_ratio/high_mean': 0.0062500000931322575, 'clip_ratio/high_max': 0.0062500000931322575, 'clip_ratio/region_mean': 0.04236111231148243, 'reward_total_mean': 0.1212349608540535, 'reward_meter_mean': 0.2789691686630249, 'reward_meter_std': 0.28043612837791443, '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_lexical_plausibility_mean': 0.8315871357917786, 'reward_lexical_plausibility_std': 0.1664452999830246, 'reward_repeat_penalty_mean': 0.7809751033782959, 'reward_repeat_penalty_std': 0.08850277960300446, 'reward_near_duplicate_penalty_mean': 0.9996334314346313, 'reward_near_duplicate_penalty_std': 0.0010368125513195992, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.2862499952316284, 'reward_judge_quality_std': 0.3945680260658264, 'reward_total_composite_mean': 0.1212349608540535, 'reward_total_composite_std': 0.2543593645095825, 'epoch': 0.03}

 19%|█▉        | 634/3300 [2:30:31<9:21:33, 12.64s/it]INFO 04-12 13:16:43 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 635/3300 [2:30:43<9:07:29, 12.33s/it]2026-04-12 13:16:55,261 | INFO | train_grpo_train | metrics_logged mode=train step=635

                                                      
{'loss': -0.0314, 'grad_norm': 6.6901655197143555, 'learning_rate': 8.07878787878788e-06, 'num_tokens': 1273144.0, 'completions/mean_length': 154.0, 'completions/min_length': 58.0, 'completions/max_length': 235.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 154.0, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 235.0, 'rewards/meter/mean': 0.9548834562301636, 'rewards/meter/std': 0.026925699785351753, 'rewards/count_adherence/mean': 0.7916666865348816, 'rewards/count_adherence/std': 0.2920915186405182, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9866071343421936, 'rewards/lexical_plausibility/std': 0.03788072615861893, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.12464234977960587, 'rewards/repeat_penalty/mean': 0.5209765434265137, 'rewards/repeat_penalty/std': 0.35144948959350586, 'rewards/near_duplicate_penalty/mean': 0.8737728595733643, 'rewards/near_duplicate_penalty/std': 0.10306568443775177, 'rewards/opening_diversity/mean': 0.9352678656578064, 'rewards/opening_diversity/std': 0.09637068212032318, 'rewards/distinct_2/mean': 0.6008220911026001, 'rewards/distinct_2/std': 0.3171120882034302, 'rewards/total_composite/mean': 0.16293153166770935, 'rewards/total_composite/std': 0.1565198451280594, 'reward': 0.16293153166770935, 'reward_std': 0.1565198451280594, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10899078100919724, 'sampling/sampling_logp_difference/max': 1.8785130977630615, 'sampling/importance_sampling_ratio/min': 0.15281717479228973, 'sampling/importance_sampling_ratio/mean': 1.0064257383346558, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8146054446697235, 'clip_ratio/low_mean': 0.06438315100967884, 'clip_ratio/low_min': 0.06438315100967884, 'clip_ratio/high_mean': 0.04923677537590265, 'clip_ratio/high_max': 0.04923677537590265, 'clip_ratio/region_mean': 0.1136199263855815, 'reward_total_mean': 0.16293153166770935, 'reward_meter_mean': 0.9548834562301636, 'reward_meter_std': 0.026925699785351753, 'reward_count_adherence_mean': 0.7916666865348816, 'reward_count_adherence_std': 0.2920915186405182, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9866071343421936, 'reward_lexical_plausibility_std': 0.03788072615861893, 'reward_repeat_penalty_mean': 0.5209765434265137, 'reward_repeat_penalty_std': 0.35144948959350586, 'reward_near_duplicate_penalty_mean': 0.8737728595733643, 'reward_near_duplicate_penalty_std': 0.10306568443775177, 'reward_opening_diversity_mean': 0.9352678656578064, 'reward_opening_diversity_std': 0.09637068212032318, 'reward_distinct_2_mean': 0.6008220911026001, 'reward_distinct_2_std': 0.3171120882034302, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.12464234977960587, 'reward_total_composite_mean': 0.16293153166770935, 'reward_total_composite_std': 0.1565198451280594, 'epoch': 0.03}

 19%|█▉        | 635/3300 [2:30:43<9:07:29, 12.33s/it]INFO 04-12 13:16:55 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 636/3300 [2:30:54<8:48:12, 11.90s/it]2026-04-12 13:17:06,158 | INFO | train_grpo_train | metrics_logged mode=train step=636

                                                      
{'loss': 0.0334, 'grad_norm': 10.770613670349121, 'learning_rate': 8.075757575757577e-06, 'num_tokens': 1276342.0, 'completions/mean_length': 171.75, 'completions/min_length': 148.0, 'completions/max_length': 187.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 171.75, 'completions/min_terminated_length': 148.0, 'completions/max_terminated_length': 187.0, 'rewards/meter/mean': 0.8720189332962036, 'rewards/meter/std': 0.15706664323806763, 'rewards/count_adherence/mean': 0.8611111640930176, 'rewards/count_adherence/std': 0.05143444612622261, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.0353553369641304, 'rewards/repeat_penalty/mean': 0.566994845867157, 'rewards/repeat_penalty/std': 0.2003743052482605, 'rewards/near_duplicate_penalty/mean': 0.9049088954925537, 'rewards/near_duplicate_penalty/std': 0.04348953068256378, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.6383098363876343, 'rewards/distinct_2/std': 0.1683952957391739, 'rewards/total_composite/mean': 0.25634610652923584, 'rewards/total_composite/std': 0.061406008899211884, 'reward': 0.25634610652923584, 'reward_std': 0.061406008899211884, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13206955790519714, 'sampling/sampling_logp_difference/max': 3.0831995010375977, 'sampling/importance_sampling_ratio/min': 0.045812446624040604, 'sampling/importance_sampling_ratio/mean': 1.006094217300415, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7453290075063705, 'clip_ratio/low_mean': 0.04905197583138943, 'clip_ratio/low_min': 0.04905197583138943, 'clip_ratio/high_mean': 0.08030822314321995, 'clip_ratio/high_max': 0.08030822314321995, 'clip_ratio/region_mean': 0.12936019897460938, 'reward_total_mean': 0.25634610652923584, 'reward_meter_mean': 0.8720189332962036, 'reward_meter_std': 0.15706664323806763, 'reward_count_adherence_mean': 0.8611111640930176, 'reward_count_adherence_std': 0.05143444612622261, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.566994845867157, 'reward_repeat_penalty_std': 0.2003743052482605, 'reward_near_duplicate_penalty_mean': 0.9049088954925537, 'reward_near_duplicate_penalty_std': 0.04348953068256378, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.6383098363876343, 'reward_distinct_2_std': 0.1683952957391739, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.0353553369641304, 'reward_total_composite_mean': 0.25634610652923584, 'reward_total_composite_std': 0.061406008899211884, 'epoch': 0.03}

 19%|█▉        | 636/3300 [2:30:54<8:48:12, 11.90s/it]INFO 04-12 13:17:06 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 637/3300 [2:31:07<9:02:12, 12.22s/it]2026-04-12 13:17:19,119 | INFO | train_grpo_train | metrics_logged mode=train step=637

                                                      
{'loss': -0.0455, 'grad_norm': 3.5608768463134766, 'learning_rate': 8.072727272727274e-06, 'num_tokens': 1277773.0, 'completions/mean_length': 157.875, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 39.833335876464844, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.38278231024742126, 'rewards/meter/std': 0.3874512016773224, '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/lexical_plausibility/mean': 0.928788423538208, 'rewards/lexical_plausibility/std': 0.11096619069576263, 'rewards/judge_quality/mean': 0.4637500047683716, 'rewards/judge_quality/std': 0.3948214054107666, 'rewards/repeat_penalty/mean': 0.9843645095825195, 'rewards/repeat_penalty/std': 0.019120698794722557, 'rewards/near_duplicate_penalty/mean': 0.9843645095825195, 'rewards/near_duplicate_penalty/std': 0.019120698794722557, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.10513200610876083, 'rewards/total_composite/std': 0.11730080097913742, 'reward': 0.10513200610876083, 'reward_std': 0.11730080097913742, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17525507509708405, 'sampling/sampling_logp_difference/max': 1.5014729499816895, 'sampling/importance_sampling_ratio/min': 0.22280174493789673, 'sampling/importance_sampling_ratio/mean': 1.0429105758666992, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9021463692188263, 'clip_ratio/low_mean': 0.0445723682641983, 'clip_ratio/low_min': 0.0445723682641983, 'clip_ratio/high_mean': 0.0713918786495924, 'clip_ratio/high_max': 0.0713918786495924, 'clip_ratio/region_mean': 0.1159642469137907, 'reward_total_mean': 0.10513200610876083, 'reward_meter_mean': 0.38278231024742126, 'reward_meter_std': 0.3874512016773224, '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_lexical_plausibility_mean': 0.928788423538208, 'reward_lexical_plausibility_std': 0.11096619069576263, 'reward_repeat_penalty_mean': 0.9843645095825195, 'reward_repeat_penalty_std': 0.019120698794722557, 'reward_near_duplicate_penalty_mean': 0.9843645095825195, 'reward_near_duplicate_penalty_std': 0.019120698794722557, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4637500047683716, 'reward_judge_quality_std': 0.3948214054107666, 'reward_total_composite_mean': 0.10513200610876083, 'reward_total_composite_std': 0.11730080097913742, 'epoch': 0.03}

 19%|█▉        | 637/3300 [2:31:07<9:02:12, 12.22s/it]INFO 04-12 13:17:19 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 638/3300 [2:31:14<8:01:38, 10.86s/it]2026-04-12 13:17:26,799 | INFO | train_grpo_train | metrics_logged mode=train step=638

                                                      
{'loss': 0.0115, 'grad_norm': 14.571146965026855, 'learning_rate': 8.069696969696971e-06, 'num_tokens': 1279259.0, 'completions/mean_length': 26.75, 'completions/min_length': 24.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.75, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.8380571603775024, 'rewards/meter/std': 0.31788626313209534, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.959099292755127, 'rewards/lexical_plausibility/std': 0.09303800761699677, 'rewards/judge_quality/mean': 0.7649999856948853, 'rewards/judge_quality/std': 0.29799333214759827, 'rewards/repeat_penalty/mean': 0.7357954382896423, 'rewards/repeat_penalty/std': 0.040176525712013245, 'rewards/near_duplicate_penalty/mean': 0.9810606241226196, 'rewards/near_duplicate_penalty/std': 0.05356868728995323, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6197974681854248, 'rewards/total_composite/std': 0.35693055391311646, 'reward': 0.6197974681854248, 'reward_std': 0.35693055391311646, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1598413586616516, 'sampling/sampling_logp_difference/max': 1.6024084091186523, 'sampling/importance_sampling_ratio/min': 0.20141084492206573, 'sampling/importance_sampling_ratio/mean': 1.0275013446807861, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9943805411458015, 'clip_ratio/low_mean': 0.06134615372866392, 'clip_ratio/low_min': 0.06134615372866392, 'clip_ratio/high_mean': 0.08046245528385043, 'clip_ratio/high_max': 0.08046245528385043, 'clip_ratio/region_mean': 0.14180860901251435, 'reward_total_mean': 0.6197974681854248, 'reward_meter_mean': 0.8380571603775024, 'reward_meter_std': 0.31788626313209534, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.959099292755127, 'reward_lexical_plausibility_std': 0.09303800761699677, 'reward_repeat_penalty_mean': 0.7357954382896423, 'reward_repeat_penalty_std': 0.040176525712013245, 'reward_near_duplicate_penalty_mean': 0.9810606241226196, 'reward_near_duplicate_penalty_std': 0.05356868728995323, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7649999856948853, 'reward_judge_quality_std': 0.29799333214759827, 'reward_total_composite_mean': 0.6197974681854248, 'reward_total_composite_std': 0.35693055391311646, 'epoch': 0.03}

 19%|█▉        | 638/3300 [2:31:14<8:01:38, 10.86s/it]INFO 04-12 13:17:27 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 639/3300 [2:31:29<8:47:33, 11.90s/it]2026-04-12 13:17:41,122 | INFO | train_grpo_train | metrics_logged mode=train step=639

                                                      
{'loss': -0.2527, 'grad_norm': 2.107557773590088, 'learning_rate': 8.066666666666667e-06, 'num_tokens': 1284153.0, 'completions/mean_length': 362.75, 'completions/min_length': 84.0, 'completions/max_length': 465.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 362.75, 'completions/min_terminated_length': 84.0, 'completions/max_terminated_length': 465.0, 'rewards/meter/mean': 0.9797854423522949, 'rewards/meter/std': 0.019569000229239464, 'rewards/count_adherence/mean': 0.8214285373687744, 'rewards/count_adherence/std': 0.2474358230829239, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.17500001192092896, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/repeat_penalty/mean': 0.10890862345695496, 'rewards/repeat_penalty/std': 0.3080401122570038, 'rewards/near_duplicate_penalty/mean': 0.6676996350288391, 'rewards/near_duplicate_penalty/std': 0.12176878750324249, 'rewards/opening_diversity/mean': 0.8863281011581421, 'rewards/opening_diversity/std': 0.09243025630712509, 'rewards/distinct_2/mean': 0.11917249113321304, 'rewards/distinct_2/std': 0.3370707035064697, 'rewards/total_composite/mean': 0.12620297074317932, 'rewards/total_composite/std': 0.022907156497240067, 'reward': 0.12620297074317932, 'reward_std': 0.022907156497240067, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.039396945387125015, 'sampling/sampling_logp_difference/max': 3.5351552963256836, 'sampling/importance_sampling_ratio/min': 0.029154228046536446, 'sampling/importance_sampling_ratio/mean': 1.0001709461212158, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.29568417742848396, 'clip_ratio/low_mean': 0.01957417675293982, 'clip_ratio/low_min': 0.01957417675293982, 'clip_ratio/high_mean': 0.026986095821484923, 'clip_ratio/high_max': 0.026986095821484923, 'clip_ratio/region_mean': 0.046560272574424744, 'reward_total_mean': 0.12620297074317932, 'reward_meter_mean': 0.9797854423522949, 'reward_meter_std': 0.019569000229239464, 'reward_count_adherence_mean': 0.8214285373687744, 'reward_count_adherence_std': 0.2474358230829239, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.10890862345695496, 'reward_repeat_penalty_std': 0.3080401122570038, 'reward_near_duplicate_penalty_mean': 0.6676996350288391, 'reward_near_duplicate_penalty_std': 0.12176878750324249, 'reward_opening_diversity_mean': 0.8863281011581421, 'reward_opening_diversity_std': 0.09243025630712509, 'reward_distinct_2_mean': 0.11917249113321304, 'reward_distinct_2_std': 0.3370707035064697, 'reward_judge_quality_mean': 0.17500001192092896, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.12620297074317932, 'reward_total_composite_std': 0.022907156497240067, 'epoch': 0.03}

 19%|█▉        | 639/3300 [2:31:29<8:47:33, 11.90s/it]INFO 04-12 13:17:41 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 640/3300 [2:31:43<9:16:48, 12.56s/it]2026-04-12 13:17:55,235 | INFO | train_grpo_train | metrics_logged mode=train step=640

                                                      
{'loss': -0.0319, 'grad_norm': 7.621056079864502, 'learning_rate': 8.063636363636364e-06, 'num_tokens': 1285615.0, 'completions/mean_length': 83.75, 'completions/min_length': 19.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 22.571430206298828, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.6029120087623596, 'rewards/meter/std': 0.49232953786849976, '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/lexical_plausibility/mean': 0.9339431524276733, 'rewards/lexical_plausibility/std': 0.12514233589172363, 'rewards/judge_quality/mean': 0.4087499976158142, 'rewards/judge_quality/std': 0.3479095697402954, 'rewards/repeat_penalty/mean': 0.7169308662414551, 'rewards/repeat_penalty/std': 0.04585881903767586, 'rewards/near_duplicate_penalty/mean': 0.9559078812599182, 'rewards/near_duplicate_penalty/std': 0.06114509701728821, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.22680015861988068, 'rewards/total_composite/std': 0.31569546461105347, 'reward': 0.22680015861988068, 'reward_std': 0.3156954348087311, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15921121835708618, 'sampling/sampling_logp_difference/max': 1.4940481185913086, 'sampling/importance_sampling_ratio/min': 0.22446216642856598, 'sampling/importance_sampling_ratio/mean': 1.0230909585952759, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8584880083799362, 'clip_ratio/low_mean': 0.0924185486510396, 'clip_ratio/low_min': 0.0924185486510396, 'clip_ratio/high_mean': 0.04479437693953514, 'clip_ratio/high_max': 0.04479437693953514, 'clip_ratio/region_mean': 0.13721292559057474, 'reward_total_mean': 0.22680015861988068, 'reward_meter_mean': 0.6029120087623596, 'reward_meter_std': 0.49232953786849976, '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_lexical_plausibility_mean': 0.9339431524276733, 'reward_lexical_plausibility_std': 0.12514233589172363, 'reward_repeat_penalty_mean': 0.7169308662414551, 'reward_repeat_penalty_std': 0.04585881903767586, 'reward_near_duplicate_penalty_mean': 0.9559078812599182, 'reward_near_duplicate_penalty_std': 0.06114509701728821, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4087499976158142, 'reward_judge_quality_std': 0.3479095697402954, 'reward_total_composite_mean': 0.22680015861988068, 'reward_total_composite_std': 0.31569546461105347, 'epoch': 0.03}

 19%|█▉        | 640/3300 [2:31:43<9:16:48, 12.56s/it]INFO 04-12 13:17:55 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 641/3300 [2:31:53<8:43:41, 11.82s/it]2026-04-12 13:18:05,320 | INFO | train_grpo_train | metrics_logged mode=train step=641

                                                      
{'loss': 0.157, 'grad_norm': 15.96769905090332, 'learning_rate': 8.060606060606061e-06, 'num_tokens': 1287435.0, 'completions/mean_length': 64.5, 'completions/min_length': 45.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.5, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.7461182475090027, 'rewards/meter/std': 0.2852329909801483, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4962499737739563, 'rewards/judge_quality/std': 0.21836650371551514, 'rewards/repeat_penalty/mean': 0.9395760297775269, 'rewards/repeat_penalty/std': 0.051755864173173904, 'rewards/near_duplicate_penalty/mean': 0.9722237586975098, 'rewards/near_duplicate_penalty/std': 0.01932915486395359, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.966033935546875, 'rewards/distinct_2/std': 0.040175117552280426, 'rewards/total_composite/mean': 0.32378101348876953, 'rewards/total_composite/std': 0.18241840600967407, 'reward': 0.32378101348876953, 'reward_std': 0.18241840600967407, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18672369420528412, 'sampling/sampling_logp_difference/max': 2.4093427658081055, 'sampling/importance_sampling_ratio/min': 0.08987434953451157, 'sampling/importance_sampling_ratio/mean': 1.025506854057312, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9243885278701782, 'clip_ratio/low_mean': 0.07249605283141136, 'clip_ratio/low_min': 0.07249605283141136, 'clip_ratio/high_mean': 0.10350418835878372, 'clip_ratio/high_max': 0.10350418835878372, 'clip_ratio/region_mean': 0.17600024119019508, 'reward_total_mean': 0.32378101348876953, 'reward_meter_mean': 0.7461182475090027, 'reward_meter_std': 0.2852329909801483, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9395760297775269, 'reward_repeat_penalty_std': 0.051755864173173904, 'reward_near_duplicate_penalty_mean': 0.9722237586975098, 'reward_near_duplicate_penalty_std': 0.01932915486395359, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.966033935546875, 'reward_distinct_2_std': 0.040175117552280426, 'reward_judge_quality_mean': 0.4962499737739563, 'reward_judge_quality_std': 0.21836650371551514, 'reward_total_composite_mean': 0.32378101348876953, 'reward_total_composite_std': 0.18241840600967407, 'epoch': 0.03}

 19%|█▉        | 641/3300 [2:31:53<8:43:41, 11.82s/it]INFO 04-12 13:18:05 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 642/3300 [2:32:03<8:26:24, 11.43s/it]2026-04-12 13:18:15,849 | INFO | train_grpo_train | metrics_logged mode=train step=642

                                                      
{'loss': 0.0134, 'grad_norm': 9.472576141357422, 'learning_rate': 8.057575757575759e-06, 'num_tokens': 1289618.0, 'completions/mean_length': 93.875, 'completions/min_length': 68.0, 'completions/max_length': 120.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 93.875, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 120.0, 'rewards/meter/mean': 0.481609046459198, 'rewards/meter/std': 0.39368680119514465, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.6929478645324707, 'rewards/repeat_penalty/std': 0.22728116810321808, 'rewards/near_duplicate_penalty/mean': 0.9295317530632019, 'rewards/near_duplicate_penalty/std': 0.04386115446686745, 'rewards/opening_diversity/mean': 0.84375, 'rewards/opening_diversity/std': 0.16366341710090637, 'rewards/distinct_2/mean': 0.8599269986152649, 'rewards/distinct_2/std': 0.10274447500705719, 'rewards/total_composite/mean': 0.16080939769744873, 'rewards/total_composite/std': 0.170904740691185, 'reward': 0.16080939769744873, 'reward_std': 0.170904740691185, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1623670756816864, 'sampling/sampling_logp_difference/max': 2.36892032623291, 'sampling/importance_sampling_ratio/min': 0.09358171373605728, 'sampling/importance_sampling_ratio/mean': 0.9873301386833191, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8296694234013557, 'clip_ratio/low_mean': 0.07854306325316429, 'clip_ratio/low_min': 0.07854306325316429, 'clip_ratio/high_mean': 0.07493629865348339, 'clip_ratio/high_max': 0.07493629865348339, 'clip_ratio/region_mean': 0.15347936190664768, 'reward_total_mean': 0.16080939769744873, 'reward_meter_mean': 0.481609046459198, 'reward_meter_std': 0.39368680119514465, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6929478645324707, 'reward_repeat_penalty_std': 0.22728116810321808, 'reward_near_duplicate_penalty_mean': 0.9295317530632019, 'reward_near_duplicate_penalty_std': 0.04386115446686745, 'reward_opening_diversity_mean': 0.84375, 'reward_opening_diversity_std': 0.16366341710090637, 'reward_distinct_2_mean': 0.8599269986152649, 'reward_distinct_2_std': 0.10274447500705719, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.16080939769744873, 'reward_total_composite_std': 0.170904740691185, 'epoch': 0.03}

 19%|█▉        | 642/3300 [2:32:03<8:26:24, 11.43s/it]INFO 04-12 13:18:16 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 643/3300 [2:32:12<7:53:33, 10.69s/it]2026-04-12 13:18:24,826 | INFO | train_grpo_train | metrics_logged mode=train step=643

                                                      
{'loss': 0.0112, 'grad_norm': 11.457493782043457, 'learning_rate': 8.054545454545454e-06, 'num_tokens': 1291562.0, 'completions/mean_length': 78.0, 'completions/min_length': 68.0, 'completions/max_length': 94.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 78.0, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.5476718544960022, 'rewards/meter/std': 0.31364360451698303, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.9320908784866333, 'rewards/repeat_penalty/std': 0.08314214646816254, 'rewards/near_duplicate_penalty/mean': 0.9673113226890564, 'rewards/near_duplicate_penalty/std': 0.027428893372416496, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9626831412315369, 'rewards/distinct_2/std': 0.07143151015043259, 'rewards/total_composite/mean': 0.10499675571918488, 'rewards/total_composite/std': 0.07011913508176804, 'reward': 0.10499675571918488, 'reward_std': 0.07011914253234863, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13167837262153625, 'sampling/sampling_logp_difference/max': 1.5653483867645264, 'sampling/importance_sampling_ratio/min': 0.24765388667583466, 'sampling/importance_sampling_ratio/mean': 1.012163758277893, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8878500610589981, 'clip_ratio/low_mean': 0.05099678970873356, 'clip_ratio/low_min': 0.05099678970873356, 'clip_ratio/high_mean': 0.06668656319379807, 'clip_ratio/high_max': 0.06668656319379807, 'clip_ratio/region_mean': 0.11768335290253162, 'reward_total_mean': 0.10499675571918488, 'reward_meter_mean': 0.5476718544960022, 'reward_meter_std': 0.31364360451698303, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9320908784866333, 'reward_repeat_penalty_std': 0.08314214646816254, 'reward_near_duplicate_penalty_mean': 0.9673113226890564, 'reward_near_duplicate_penalty_std': 0.027428893372416496, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9626831412315369, 'reward_distinct_2_std': 0.07143151015043259, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.10499675571918488, 'reward_total_composite_std': 0.07011913508176804, 'epoch': 0.03}

 19%|█▉        | 643/3300 [2:32:12<7:53:33, 10.69s/it]INFO 04-12 13:18:25 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 644/3300 [2:32:21<7:31:15, 10.19s/it]2026-04-12 13:18:33,847 | INFO | train_grpo_train | metrics_logged mode=train step=644

                                                      
{'loss': 0.074, 'grad_norm': 14.46362018585205, 'learning_rate': 8.051515151515153e-06, 'num_tokens': 1293507.0, 'completions/mean_length': 49.125, 'completions/min_length': 41.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.125, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.6963322162628174, 'rewards/meter/std': 0.3832204043865204, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.975322425365448, 'rewards/lexical_plausibility/std': 0.04015253856778145, 'rewards/judge_quality/mean': 0.8524999618530273, 'rewards/judge_quality/std': 0.20331187546253204, 'rewards/repeat_penalty/mean': 0.9885950684547424, 'rewards/repeat_penalty/std': 0.014315770007669926, 'rewards/near_duplicate_penalty/mean': 0.9885950684547424, 'rewards/near_duplicate_penalty/std': 0.014315770007669926, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5737149715423584, 'rewards/total_composite/std': 0.3512689769268036, 'reward': 0.5737149715423584, 'reward_std': 0.3512689769268036, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14565438032150269, 'sampling/sampling_logp_difference/max': 1.4939188957214355, 'sampling/importance_sampling_ratio/min': 0.2244911789894104, 'sampling/importance_sampling_ratio/mean': 1.0124223232269287, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9569375514984131, 'clip_ratio/low_mean': 0.09816683270037174, 'clip_ratio/low_min': 0.09816683270037174, 'clip_ratio/high_mean': 0.05961468303576112, 'clip_ratio/high_max': 0.05961468303576112, 'clip_ratio/region_mean': 0.15778151573613286, 'reward_total_mean': 0.5737149715423584, 'reward_meter_mean': 0.6963322162628174, 'reward_meter_std': 0.3832204043865204, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.975322425365448, 'reward_lexical_plausibility_std': 0.04015253856778145, 'reward_repeat_penalty_mean': 0.9885950684547424, 'reward_repeat_penalty_std': 0.014315770007669926, 'reward_near_duplicate_penalty_mean': 0.9885950684547424, 'reward_near_duplicate_penalty_std': 0.014315770007669926, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8524999618530273, 'reward_judge_quality_std': 0.20331187546253204, 'reward_total_composite_mean': 0.5737149715423584, 'reward_total_composite_std': 0.3512689769268036, 'epoch': 0.03}

 20%|█▉        | 644/3300 [2:32:21<7:31:15, 10.19s/it]INFO 04-12 13:18:34 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 645/3300 [2:32:36<8:35:20, 11.65s/it]2026-04-12 13:18:48,881 | INFO | train_grpo_train | metrics_logged mode=train step=645

                                                      
{'loss': -0.0593, 'grad_norm': 5.030296325683594, 'learning_rate': 8.048484848484849e-06, 'num_tokens': 1295481.0, 'completions/mean_length': 130.75, 'completions/min_length': 71.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 76.28572082519531, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.45199140906333923, 'rewards/meter/std': 0.33480334281921387, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9682062864303589, 'rewards/lexical_plausibility/std': 0.08992617577314377, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.1414213478565216, 'rewards/repeat_penalty/mean': 0.9182196855545044, 'rewards/repeat_penalty/std': 0.06554044783115387, 'rewards/near_duplicate_penalty/mean': 0.9543733596801758, 'rewards/near_duplicate_penalty/std': 0.027052002027630806, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9612371921539307, 'rewards/distinct_2/std': 0.046426400542259216, 'rewards/total_composite/mean': 0.08623696863651276, 'rewards/total_composite/std': 0.07367822527885437, 'reward': 0.08623696863651276, 'reward_std': 0.07367822527885437, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1614648401737213, 'sampling/sampling_logp_difference/max': 1.3377704620361328, 'sampling/importance_sampling_ratio/min': 0.2624301314353943, 'sampling/importance_sampling_ratio/mean': 1.030680537223816, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2282000184059143, 'clip_ratio/low_mean': 0.0690268985927105, 'clip_ratio/low_min': 0.0690268985927105, 'clip_ratio/high_mean': 0.07994715869426727, 'clip_ratio/high_max': 0.07994715869426727, 'clip_ratio/region_mean': 0.14897405728697777, 'reward_total_mean': 0.08623696863651276, 'reward_meter_mean': 0.45199140906333923, 'reward_meter_std': 0.33480334281921387, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9682062864303589, 'reward_lexical_plausibility_std': 0.08992617577314377, 'reward_repeat_penalty_mean': 0.9182196855545044, 'reward_repeat_penalty_std': 0.06554044783115387, 'reward_near_duplicate_penalty_mean': 0.9543733596801758, 'reward_near_duplicate_penalty_std': 0.027052002027630806, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9612371921539307, 'reward_distinct_2_std': 0.046426400542259216, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.1414213478565216, 'reward_total_composite_mean': 0.08623696863651276, 'reward_total_composite_std': 0.07367822527885437, 'epoch': 0.03}

 20%|█▉        | 645/3300 [2:32:36<8:35:20, 11.65s/it]INFO 04-12 13:18:49 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 646/3300 [2:32:46<8:07:30, 11.02s/it]2026-04-12 13:18:58,445 | INFO | train_grpo_train | metrics_logged mode=train step=646

                                                      
{'loss': 0.0588, 'grad_norm': 17.052248001098633, 'learning_rate': 8.045454545454546e-06, 'num_tokens': 1297184.0, 'completions/mean_length': 47.875, 'completions/min_length': 43.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.875, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.6372324228286743, 'rewards/meter/std': 0.37157922983169556, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5174999833106995, 'rewards/judge_quality/std': 0.26868730783462524, 'rewards/repeat_penalty/mean': 0.9930521249771118, 'rewards/repeat_penalty/std': 0.0094807930290699, 'rewards/near_duplicate_penalty/mean': 0.9930521249771118, 'rewards/near_duplicate_penalty/std': 0.0094807930290699, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3884514570236206, 'rewards/total_composite/std': 0.34655410051345825, 'reward': 0.3884514570236206, 'reward_std': 0.34655407071113586, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15788908302783966, 'sampling/sampling_logp_difference/max': 2.327352523803711, 'sampling/importance_sampling_ratio/min': 0.18768973648548126, 'sampling/importance_sampling_ratio/mean': 1.0095030069351196, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.687575925141573, 'clip_ratio/low_mean': 0.07169033400714397, 'clip_ratio/low_min': 0.07169033400714397, 'clip_ratio/high_mean': 0.060691287741065025, 'clip_ratio/high_max': 0.060691287741065025, 'clip_ratio/region_mean': 0.132381621748209, 'reward_total_mean': 0.3884514570236206, 'reward_meter_mean': 0.6372324228286743, 'reward_meter_std': 0.37157922983169556, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9930521249771118, 'reward_repeat_penalty_std': 0.0094807930290699, 'reward_near_duplicate_penalty_mean': 0.9930521249771118, 'reward_near_duplicate_penalty_std': 0.0094807930290699, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5174999833106995, 'reward_judge_quality_std': 0.26868730783462524, 'reward_total_composite_mean': 0.3884514570236206, 'reward_total_composite_std': 0.34655410051345825, 'epoch': 0.03}

 20%|█▉        | 646/3300 [2:32:46<8:07:30, 11.02s/it]INFO 04-12 13:18:58 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 647/3300 [2:32:55<7:44:38, 10.51s/it]2026-04-12 13:19:07,755 | INFO | train_grpo_train | metrics_logged mode=train step=647

                                                      
{'loss': 0.0042, 'grad_norm': 19.14881706237793, 'learning_rate': 8.042424242424243e-06, 'num_tokens': 1298598.0, 'completions/mean_length': 25.75, 'completions/min_length': 23.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.75, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.5218876600265503, 'rewards/meter/std': 0.49559152126312256, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.966911792755127, 'rewards/lexical_plausibility/std': 0.09358766674995422, 'rewards/judge_quality/mean': 0.6425000429153442, 'rewards/judge_quality/std': 0.4080178737640381, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.19465665519237518, 'rewards/total_composite/std': 0.28726205229759216, 'reward': 0.19465665519237518, 'reward_std': 0.28726205229759216, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11895209550857544, 'sampling/sampling_logp_difference/max': 1.2821241617202759, 'sampling/importance_sampling_ratio/min': 0.27744731307029724, 'sampling/importance_sampling_ratio/mean': 1.0020582675933838, 'sampling/importance_sampling_ratio/max': 1.704353928565979, 'entropy': 0.8862093053758144, 'clip_ratio/low_mean': 0.09759674221277237, 'clip_ratio/low_min': 0.09759674221277237, 'clip_ratio/high_mean': 0.014423076994717121, 'clip_ratio/high_max': 0.014423076994717121, 'clip_ratio/region_mean': 0.11201981920748949, 'reward_total_mean': 0.19465665519237518, 'reward_meter_mean': 0.5218876600265503, 'reward_meter_std': 0.49559152126312256, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.966911792755127, 'reward_lexical_plausibility_std': 0.09358766674995422, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6425000429153442, 'reward_judge_quality_std': 0.4080178737640381, 'reward_total_composite_mean': 0.19465665519237518, 'reward_total_composite_std': 0.28726205229759216, 'epoch': 0.03}

 20%|█▉        | 647/3300 [2:32:55<7:44:38, 10.51s/it]INFO 04-12 13:19:07 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 648/3300 [2:33:10<8:44:48, 11.87s/it]2026-04-12 13:19:22,817 | INFO | train_grpo_train | metrics_logged mode=train step=648

                                                      
{'loss': -0.1016, 'grad_norm': 3.156404733657837, 'learning_rate': 8.03939393939394e-06, 'num_tokens': 1300181.0, 'completions/mean_length': 165.875, 'completions/min_length': 47.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 50.5, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.749919056892395, 'rewards/meter/std': 0.3598979711532593, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9536298513412476, 'rewards/lexical_plausibility/std': 0.08603141456842422, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.155264750123024, 'rewards/repeat_penalty/mean': 0.9071629643440247, 'rewards/repeat_penalty/std': 0.12951858341693878, 'rewards/near_duplicate_penalty/mean': 0.9722334742546082, 'rewards/near_duplicate_penalty/std': 0.03392884507775307, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9598901271820068, 'rewards/distinct_2/std': 0.05960153788328171, 'rewards/total_composite/mean': 0.23855365812778473, 'rewards/total_composite/std': 0.17867974936962128, 'reward': 0.23855365812778473, 'reward_std': 0.17867976427078247, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17387282848358154, 'sampling/sampling_logp_difference/max': 2.792581081390381, 'sampling/importance_sampling_ratio/min': 0.06126288324594498, 'sampling/importance_sampling_ratio/mean': 1.0311232805252075, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.774294950067997, 'clip_ratio/low_mean': 0.01595744676887989, 'clip_ratio/low_min': 0.01595744676887989, 'clip_ratio/high_mean': 0.09739032946527004, 'clip_ratio/high_max': 0.09739032946527004, 'clip_ratio/region_mean': 0.11334777623414993, 'reward_total_mean': 0.23855365812778473, 'reward_meter_mean': 0.749919056892395, 'reward_meter_std': 0.3598979711532593, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9536298513412476, 'reward_lexical_plausibility_std': 0.08603141456842422, 'reward_repeat_penalty_mean': 0.9071629643440247, 'reward_repeat_penalty_std': 0.12951858341693878, 'reward_near_duplicate_penalty_mean': 0.9722334742546082, 'reward_near_duplicate_penalty_std': 0.03392884507775307, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9598901271820068, 'reward_distinct_2_std': 0.05960153788328171, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.155264750123024, 'reward_total_composite_mean': 0.23855365812778473, 'reward_total_composite_std': 0.17867974936962128, 'epoch': 0.03}

 20%|█▉        | 648/3300 [2:33:10<8:44:48, 11.87s/it]INFO 04-12 13:19:23 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 649/3300 [2:33:18<7:52:05, 10.68s/it]2026-04-12 13:19:30,728 | INFO | train_grpo_train | metrics_logged mode=train step=649

                                                      
{'loss': 0.0213, 'grad_norm': 18.302217483520508, 'learning_rate': 8.036363636363636e-06, 'num_tokens': 1301648.0, 'completions/mean_length': 24.375, 'completions/min_length': 22.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.375, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.9763913154602051, 'rewards/meter/std': 0.014494826085865498, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9574281573295593, 'rewards/lexical_plausibility/std': 0.09763853251934052, 'rewards/judge_quality/mean': 0.9387500286102295, 'rewards/judge_quality/std': 0.015526460483670235, 'rewards/repeat_penalty/mean': 0.7201539278030396, 'rewards/repeat_penalty/std': 0.052037425339221954, 'rewards/near_duplicate_penalty/mean': 0.9602053165435791, 'rewards/near_duplicate_penalty/std': 0.06938322633504868, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.9167482852935791, 'rewards/total_composite/std': 0.027317216619849205, 'reward': 0.9167482852935791, 'reward_std': 0.027317220345139503, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1170336902141571, 'sampling/sampling_logp_difference/max': 1.5248045921325684, 'sampling/importance_sampling_ratio/min': 0.21766360104084015, 'sampling/importance_sampling_ratio/mean': 0.9827519655227661, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6033441759645939, 'clip_ratio/low_mean': 0.04053758829832077, 'clip_ratio/low_min': 0.04053758829832077, 'clip_ratio/high_mean': 0.09080864861607552, 'clip_ratio/high_max': 0.09080864861607552, 'clip_ratio/region_mean': 0.13134623691439629, 'reward_total_mean': 0.9167482852935791, 'reward_meter_mean': 0.9763913154602051, 'reward_meter_std': 0.014494826085865498, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9574281573295593, 'reward_lexical_plausibility_std': 0.09763853251934052, 'reward_repeat_penalty_mean': 0.7201539278030396, 'reward_repeat_penalty_std': 0.052037425339221954, 'reward_near_duplicate_penalty_mean': 0.9602053165435791, 'reward_near_duplicate_penalty_std': 0.06938322633504868, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.9387500286102295, 'reward_judge_quality_std': 0.015526460483670235, 'reward_total_composite_mean': 0.9167482852935791, 'reward_total_composite_std': 0.027317216619849205, 'epoch': 0.03}

 20%|█▉        | 649/3300 [2:33:18<7:52:05, 10.68s/it]INFO 04-12 13:19:30 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 650/3300 [2:33:26<7:20:22,  9.97s/it]2026-04-12 13:19:39,031 | INFO | train_grpo_train | metrics_logged mode=train step=650

                                                      
{'loss': 0.0548, 'grad_norm': 87.50554656982422, 'learning_rate': 8.033333333333335e-06, 'num_tokens': 1303056.0, 'completions/mean_length': 28.0, 'completions/min_length': 23.0, 'completions/max_length': 31.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.0, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.6164331436157227, 'rewards/meter/std': 0.44052690267562866, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5425000190734863, 'rewards/judge_quality/std': 0.34087073802948, 'rewards/repeat_penalty/mean': 0.7469173073768616, 'rewards/repeat_penalty/std': 0.00871915090829134, 'rewards/near_duplicate_penalty/mean': 0.9958897829055786, 'rewards/near_duplicate_penalty/std': 0.011625555343925953, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2809414267539978, 'rewards/total_composite/std': 0.30079224705696106, 'reward': 0.2809414267539978, 'reward_std': 0.30079224705696106, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.177694171667099, 'sampling/sampling_logp_difference/max': 2.093334197998047, 'sampling/importance_sampling_ratio/min': 0.21737733483314514, 'sampling/importance_sampling_ratio/mean': 1.0162487030029297, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.173734374344349, 'clip_ratio/low_mean': 0.053314114455133677, 'clip_ratio/low_min': 0.053314114455133677, 'clip_ratio/high_mean': 0.06503349542617798, 'clip_ratio/high_max': 0.06503349542617798, 'clip_ratio/region_mean': 0.11834760988131166, 'reward_total_mean': 0.2809414267539978, 'reward_meter_mean': 0.6164331436157227, 'reward_meter_std': 0.44052690267562866, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7469173073768616, 'reward_repeat_penalty_std': 0.00871915090829134, 'reward_near_duplicate_penalty_mean': 0.9958897829055786, 'reward_near_duplicate_penalty_std': 0.011625555343925953, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5425000190734863, 'reward_judge_quality_std': 0.34087073802948, 'reward_total_composite_mean': 0.2809414267539978, 'reward_total_composite_std': 0.30079224705696106, 'epoch': 0.03}

 20%|█▉        | 650/3300 [2:33:27<7:20:22,  9.97s/it]INFO 04-12 13:19:39 [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 [00:11<01:04,  5.90s/it]

 23%|██▎       | 3/13 [00:19<01:06,  6.62s/it]

 31%|███       | 4/13 [00:30<01:14,  8.33s/it]

 38%|███▊      | 5/13 [00:42<01:16,  9.53s/it]

 46%|████▌     | 6/13 [00:54<01:11, 10.22s/it]

 54%|█████▍    | 7/13 [01:02<00:58,  9.77s/it]

 62%|██████▏   | 8/13 [01:14<00:51, 10.32s/it]

 69%|██████▉   | 9/13 [01:25<00:42, 10.70s/it]

 77%|███████▋  | 10/13 [01:36<00:32, 10.75s/it]

 85%|████████▍ | 11/13 [01:48<00:22, 11.14s/it]

 92%|█████████▏| 12/13 [01:59<00:11, 11.14s/it]

100%|██████████| 13/13 [02:11<00:00, 11.13s/it]2026-04-12 13:22:02,242 | INFO | train_grpo_train | metrics_logged mode=eval step=650

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 143.1603, 'eval_samples_per_second': 0.726, 'eval_steps_per_second': 0.091, 'eval_num_tokens': 1303056.0, 'eval_completions/mean_length': 186.21153846153845, 'eval_completions/min_length': 44.46153846153846, 'eval_completions/max_length': 434.9230769230769, 'eval_completions/clipped_ratio': 0.04807692307692308, 'eval_completions/mean_terminated_length': 169.38553443321814, 'eval_completions/min_terminated_length': 44.46153846153846, 'eval_completions/max_terminated_length': 405.0769230769231, 'eval_rewards/meter/mean': 0.5176326403251061, 'eval_rewards/meter/std': 0.38338013337208676, 'eval_rewards/count_adherence/mean': 0.9161025790067819, 'eval_rewards/count_adherence/std': 0.10158824261564475, 'eval_rewards/arabic_clean/mean': 0.9134615384615384, 'eval_rewards/arabic_clean/std': 0.22598336522395795, 'eval_rewards/lexical_plausibility/mean': 0.9886978818820074, 'eval_rewards/lexical_plausibility/std': 0.03196719833291494, 'eval_rewards/judge_quality/mean': 0.32038461245023286, 'eval_rewards/judge_quality/std': 0.18748051501237428, 'eval_rewards/repeat_penalty/mean': 0.6313059673859522, 'eval_rewards/repeat_penalty/std': 0.3831306535464067, 'eval_rewards/near_duplicate_penalty/mean': 0.882120247070606, 'eval_rewards/near_duplicate_penalty/std': 0.14898592720811182, 'eval_rewards/opening_diversity/mean': 0.935858442233159, 'eval_rewards/opening_diversity/std': 0.13979974389076233, 'eval_rewards/distinct_2/mean': 0.6888675918945899, 'eval_rewards/distinct_2/std': 0.37871775661523527, 'eval_rewards/total_composite/mean': 0.15516162205200928, 'eval_rewards/total_composite/std': 0.1773617290533506, 'eval_reward': 0.15516162205200928, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.049533164558502346, 'eval_sampling/sampling_logp_difference/max': 1.100408700796274, 'eval_sampling/importance_sampling_ratio/min': 0.33582237362861633, 'eval_sampling/importance_sampling_ratio/mean': 1.0101247934194713, 'eval_sampling/importance_sampling_ratio/max': 1.5279319011248076, 'eval_entropy': 0.5434640416732202, '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': 0.15516162205200928, 'eval_reward_meter_mean': 0.5176326403251061, 'eval_reward_meter_std': 0.38338013337208676, 'eval_reward_count_adherence_mean': 0.9161025790067819, 'eval_reward_count_adherence_std': 0.10158824261564475, 'eval_reward_arabic_clean_mean': 0.9134615384615384, 'eval_reward_arabic_clean_std': 0.22598336522395795, 'eval_reward_lexical_plausibility_mean': 0.9886978818820074, 'eval_reward_lexical_plausibility_std': 0.03196719833291494, 'eval_reward_repeat_penalty_mean': 0.6313059673859522, 'eval_reward_repeat_penalty_std': 0.3831306535464067, 'eval_reward_near_duplicate_penalty_mean': 0.882120247070606, 'eval_reward_near_duplicate_penalty_std': 0.14898592720811182, 'eval_reward_opening_diversity_mean': 0.935858442233159, 'eval_reward_opening_diversity_std': 0.13979974389076233, 'eval_reward_distinct_2_mean': 0.6888675918945899, 'eval_reward_distinct_2_std': 0.37871775661523527, 'eval_reward_judge_quality_mean': 0.32038461245023286, 'eval_reward_judge_quality_std': 0.18748051501237428, 'eval_reward_total_composite_mean': 0.15516162205200928, 'eval_reward_total_composite_std': 0.1773617290533506, 'epoch': 0.03}

 20%|█▉        | 650/3300 [2:35:50<7:20:22,  9.97s/it]

100%|██████████| 13/13 [02:11<00:00, 11.13s/it]

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

 20%|█▉        | 651/3300 [2:36:01<39:13:40, 53.31s/it]2026-04-12 13:22:13,472 | INFO | train_grpo_train | metrics_logged mode=train step=651

                                                       
{'loss': 0.0201, 'grad_norm': 18.738645553588867, 'learning_rate': 8.03030303030303e-06, 'num_tokens': 1304488.0, 'completions/mean_length': 26.0, 'completions/min_length': 22.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.0, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.9726604223251343, 'rewards/meter/std': 0.030426910147070885, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6349999904632568, 'rewards/judge_quality/std': 0.3184336721897125, 'rewards/repeat_penalty/mean': 0.7372881174087524, 'rewards/repeat_penalty/std': 0.025295833125710487, 'rewards/near_duplicate_penalty/mean': 0.9830508232116699, 'rewards/near_duplicate_penalty/std': 0.033727772533893585, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6145163774490356, 'rewards/total_composite/std': 0.30505234003067017, 'reward': 0.6145163774490356, 'reward_std': 0.30505234003067017, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13733744621276855, 'sampling/sampling_logp_difference/max': 1.0213537216186523, 'sampling/importance_sampling_ratio/min': 0.3601071536540985, 'sampling/importance_sampling_ratio/mean': 1.0433086156845093, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8922332152724266, 'clip_ratio/low_mean': 0.06401891633868217, 'clip_ratio/low_min': 0.06401891633868217, 'clip_ratio/high_mean': 0.06468990864232183, 'clip_ratio/high_max': 0.06468990864232183, 'clip_ratio/region_mean': 0.128708824981004, 'reward_total_mean': 0.6145163774490356, 'reward_meter_mean': 0.9726604223251343, 'reward_meter_std': 0.030426910147070885, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7372881174087524, 'reward_repeat_penalty_std': 0.025295833125710487, 'reward_near_duplicate_penalty_mean': 0.9830508232116699, 'reward_near_duplicate_penalty_std': 0.033727772533893585, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6349999904632568, 'reward_judge_quality_std': 0.3184336721897125, 'reward_total_composite_mean': 0.6145163774490356, 'reward_total_composite_std': 0.30505234003067017, 'epoch': 0.03}

 20%|█▉        | 651/3300 [2:36:01<39:13:40, 53.31s/it]INFO 04-12 13:22:13 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 652/3300 [2:36:12<29:59:30, 40.77s/it]2026-04-12 13:22:24,991 | INFO | train_grpo_train | metrics_logged mode=train step=652

                                                       
{'loss': 0.0465, 'grad_norm': 5.845699310302734, 'learning_rate': 8.027272727272728e-06, 'num_tokens': 1308269.0, 'completions/mean_length': 258.625, 'completions/min_length': 224.0, 'completions/max_length': 287.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 258.625, 'completions/min_terminated_length': 224.0, 'completions/max_terminated_length': 287.0, 'rewards/meter/mean': 0.8591511249542236, 'rewards/meter/std': 0.3451387882232666, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.1414213478565216, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 0.488717257976532, 'rewards/near_duplicate_penalty/std': 0.11153869330883026, 'rewards/opening_diversity/mean': 0.7508928775787354, 'rewards/opening_diversity/std': 0.25745269656181335, 'rewards/distinct_2/mean': 0.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.12882253527641296, 'rewards/total_composite/std': 0.0519123338162899, 'reward': 0.12882253527641296, 'reward_std': 0.0519123300909996, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03103749454021454, 'sampling/sampling_logp_difference/max': 1.6186375617980957, 'sampling/importance_sampling_ratio/min': 0.19816851615905762, 'sampling/importance_sampling_ratio/mean': 1.0036224126815796, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.18352647498250008, 'clip_ratio/low_mean': 0.0030487803742289543, 'clip_ratio/low_min': 0.0030487803742289543, 'clip_ratio/high_mean': 0.026172100100666285, 'clip_ratio/high_max': 0.026172100100666285, 'clip_ratio/region_mean': 0.02922088047489524, 'reward_total_mean': 0.12882253527641296, 'reward_meter_mean': 0.8591511249542236, 'reward_meter_std': 0.3451387882232666, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.1414213478565216, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.0, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 0.488717257976532, 'reward_near_duplicate_penalty_std': 0.11153869330883026, 'reward_opening_diversity_mean': 0.7508928775787354, 'reward_opening_diversity_std': 0.25745269656181335, 'reward_distinct_2_mean': 0.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.12882253527641296, 'reward_total_composite_std': 0.0519123338162899, 'epoch': 0.03}

 20%|█▉        | 652/3300 [2:36:12<29:59:30, 40.77s/it]INFO 04-12 13:22:25 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 653/3300 [2:36:22<23:01:47, 31.32s/it]2026-04-12 13:22:34,256 | INFO | train_grpo_train | metrics_logged mode=train step=653

                                                       
{'loss': -0.0468, 'grad_norm': 14.32152271270752, 'learning_rate': 8.024242424242425e-06, 'num_tokens': 1309991.0, 'completions/mean_length': 51.25, 'completions/min_length': 27.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.25, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.767805278301239, 'rewards/meter/std': 0.29257190227508545, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4462500214576721, 'rewards/judge_quality/std': 0.2173829823732376, 'rewards/repeat_penalty/mean': 0.9154626131057739, 'rewards/repeat_penalty/std': 0.09929834306240082, 'rewards/near_duplicate_penalty/mean': 0.9682186841964722, 'rewards/near_duplicate_penalty/std': 0.023648826405405998, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9752647876739502, 'rewards/distinct_2/std': 0.034148700535297394, 'rewards/total_composite/mean': 0.36292827129364014, 'rewards/total_composite/std': 0.22661900520324707, 'reward': 0.36292827129364014, 'reward_std': 0.22661900520324707, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1530527025461197, 'sampling/sampling_logp_difference/max': 1.8436579704284668, 'sampling/importance_sampling_ratio/min': 0.1582375317811966, 'sampling/importance_sampling_ratio/mean': 1.0349364280700684, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9279947057366371, 'clip_ratio/low_mean': 0.09473566338419914, 'clip_ratio/low_min': 0.09473566338419914, 'clip_ratio/high_mean': 0.06286514457315207, 'clip_ratio/high_max': 0.06286514457315207, 'clip_ratio/region_mean': 0.1576008079573512, 'reward_total_mean': 0.36292827129364014, 'reward_meter_mean': 0.767805278301239, 'reward_meter_std': 0.29257190227508545, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9154626131057739, 'reward_repeat_penalty_std': 0.09929834306240082, 'reward_near_duplicate_penalty_mean': 0.9682186841964722, 'reward_near_duplicate_penalty_std': 0.023648826405405998, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9752647876739502, 'reward_distinct_2_std': 0.034148700535297394, 'reward_judge_quality_mean': 0.4462500214576721, 'reward_judge_quality_std': 0.2173829823732376, 'reward_total_composite_mean': 0.36292827129364014, 'reward_total_composite_std': 0.22661900520324707, 'epoch': 0.03}

 20%|█▉        | 653/3300 [2:36:22<23:01:47, 31.32s/it]INFO 04-12 13:22:34 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 654/3300 [2:36:31<18:13:27, 24.80s/it]2026-04-12 13:22:43,824 | INFO | train_grpo_train | metrics_logged mode=train step=654

                                                       
{'loss': 0.0469, 'grad_norm': 10.498310089111328, 'learning_rate': 8.021212121212122e-06, 'num_tokens': 1312341.0, 'completions/mean_length': 106.75, 'completions/min_length': 97.0, 'completions/max_length': 118.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 106.75, 'completions/min_terminated_length': 97.0, 'completions/max_terminated_length': 118.0, 'rewards/meter/mean': 0.9362326264381409, 'rewards/meter/std': 0.09628015756607056, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.36249998211860657, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.7272390723228455, 'rewards/repeat_penalty/std': 0.16425026953220367, 'rewards/near_duplicate_penalty/mean': 0.9148551821708679, 'rewards/near_duplicate_penalty/std': 0.04300973564386368, 'rewards/opening_diversity/mean': 0.984375, 'rewards/opening_diversity/std': 0.0289318785071373, 'rewards/distinct_2/mean': 0.8128030300140381, 'rewards/distinct_2/std': 0.12252521514892578, 'rewards/total_composite/mean': 0.34084993600845337, 'rewards/total_composite/std': 0.10532856732606888, 'reward': 0.34084993600845337, 'reward_std': 0.10532856732606888, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14149324595928192, 'sampling/sampling_logp_difference/max': 1.6551990509033203, 'sampling/importance_sampling_ratio/min': 0.19105401635169983, 'sampling/importance_sampling_ratio/mean': 1.003642201423645, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7534263432025909, 'clip_ratio/low_mean': 0.02264983393251896, 'clip_ratio/low_min': 0.02264983393251896, 'clip_ratio/high_mean': 0.11159832589328289, 'clip_ratio/high_max': 0.11159832589328289, 'clip_ratio/region_mean': 0.13424815982580185, 'reward_total_mean': 0.34084993600845337, 'reward_meter_mean': 0.9362326264381409, 'reward_meter_std': 0.09628015756607056, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7272390723228455, 'reward_repeat_penalty_std': 0.16425026953220367, 'reward_near_duplicate_penalty_mean': 0.9148551821708679, 'reward_near_duplicate_penalty_std': 0.04300973564386368, 'reward_opening_diversity_mean': 0.984375, 'reward_opening_diversity_std': 0.0289318785071373, 'reward_distinct_2_mean': 0.8128030300140381, 'reward_distinct_2_std': 0.12252521514892578, 'reward_judge_quality_mean': 0.36249998211860657, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.34084993600845337, 'reward_total_composite_std': 0.10532856732606888, 'epoch': 0.03}

 20%|█▉        | 654/3300 [2:36:31<18:13:27, 24.80s/it]INFO 04-12 13:22:44 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 655/3300 [2:36:46<15:58:45, 21.75s/it]2026-04-12 13:22:58,465 | INFO | train_grpo_train | metrics_logged mode=train step=655

                                                       
{'loss': -0.0229, 'grad_norm': 6.326838970184326, 'learning_rate': 8.018181818181818e-06, 'num_tokens': 1313915.0, 'completions/mean_length': 102.75, 'completions/min_length': 41.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 44.28571701049805, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.3526526093482971, 'rewards/meter/std': 0.3082861304283142, '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/lexical_plausibility/mean': 0.9672703742980957, 'rewards/lexical_plausibility/std': 0.09257334470748901, 'rewards/judge_quality/mean': 0.4337499737739563, 'rewards/judge_quality/std': 0.23868314921855927, 'rewards/repeat_penalty/mean': 0.9533663392066956, 'rewards/repeat_penalty/std': 0.04521464928984642, 'rewards/near_duplicate_penalty/mean': 0.9627457857131958, 'rewards/near_duplicate_penalty/std': 0.02561154030263424, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.989881157875061, 'rewards/distinct_2/std': 0.027029765769839287, 'rewards/total_composite/mean': 0.14078126847743988, 'rewards/total_composite/std': 0.1517435908317566, 'reward': 0.14078126847743988, 'reward_std': 0.15174360573291779, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15908733010292053, 'sampling/sampling_logp_difference/max': 1.8897331953048706, 'sampling/importance_sampling_ratio/min': 0.1511121243238449, 'sampling/importance_sampling_ratio/mean': 1.0027285814285278, 'sampling/importance_sampling_ratio/max': 1.9810218811035156, 'entropy': 0.7259085550904274, 'clip_ratio/low_mean': 0.051550510339438915, 'clip_ratio/low_min': 0.051550510339438915, 'clip_ratio/high_mean': 0.0810845885425806, 'clip_ratio/high_max': 0.0810845885425806, 'clip_ratio/region_mean': 0.13263509888201952, 'reward_total_mean': 0.14078126847743988, 'reward_meter_mean': 0.3526526093482971, 'reward_meter_std': 0.3082861304283142, '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_lexical_plausibility_mean': 0.9672703742980957, 'reward_lexical_plausibility_std': 0.09257334470748901, 'reward_repeat_penalty_mean': 0.9533663392066956, 'reward_repeat_penalty_std': 0.04521464928984642, 'reward_near_duplicate_penalty_mean': 0.9627457857131958, 'reward_near_duplicate_penalty_std': 0.02561154030263424, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.989881157875061, 'reward_distinct_2_std': 0.027029765769839287, 'reward_judge_quality_mean': 0.4337499737739563, 'reward_judge_quality_std': 0.23868314921855927, 'reward_total_composite_mean': 0.14078126847743988, 'reward_total_composite_std': 0.1517435908317566, 'epoch': 0.03}

 20%|█▉        | 655/3300 [2:36:46<15:58:45, 21.75s/it]INFO 04-12 13:22:58 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 656/3300 [2:36:54<13:02:49, 17.76s/it]2026-04-12 13:23:06,931 | INFO | train_grpo_train | metrics_logged mode=train step=656

                                                       
{'loss': 0.064, 'grad_norm': 13.405967712402344, 'learning_rate': 8.015151515151515e-06, 'num_tokens': 1315539.0, 'completions/mean_length': 45.0, 'completions/min_length': 38.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.0, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.48697662353515625, 'rewards/meter/std': 0.3975958228111267, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.39625000953674316, 'rewards/judge_quality/std': 0.27458736300468445, 'rewards/repeat_penalty/mean': 0.9424288272857666, 'rewards/repeat_penalty/std': 0.07438904047012329, 'rewards/near_duplicate_penalty/mean': 0.9495924711227417, 'rewards/near_duplicate_penalty/std': 0.05471259355545044, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9912587404251099, 'rewards/distinct_2/std': 0.024724015966057777, 'rewards/total_composite/mean': 0.18401092290878296, 'rewards/total_composite/std': 0.21311181783676147, 'reward': 0.18401092290878296, 'reward_std': 0.21311181783676147, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14915597438812256, 'sampling/sampling_logp_difference/max': 1.995527982711792, 'sampling/importance_sampling_ratio/min': 0.13594186305999756, 'sampling/importance_sampling_ratio/mean': 0.9948115944862366, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8379657715559006, 'clip_ratio/low_mean': 0.07851754873991013, 'clip_ratio/low_min': 0.07851754873991013, 'clip_ratio/high_mean': 0.05291762016713619, 'clip_ratio/high_max': 0.05291762016713619, 'clip_ratio/region_mean': 0.13143516890704632, 'reward_total_mean': 0.18401092290878296, 'reward_meter_mean': 0.48697662353515625, 'reward_meter_std': 0.3975958228111267, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9424288272857666, 'reward_repeat_penalty_std': 0.07438904047012329, 'reward_near_duplicate_penalty_mean': 0.9495924711227417, 'reward_near_duplicate_penalty_std': 0.05471259355545044, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9912587404251099, 'reward_distinct_2_std': 0.024724015966057777, 'reward_judge_quality_mean': 0.39625000953674316, 'reward_judge_quality_std': 0.27458736300468445, 'reward_total_composite_mean': 0.18401092290878296, 'reward_total_composite_std': 0.21311181783676147, 'epoch': 0.03}

 20%|█▉        | 656/3300 [2:36:54<13:02:49, 17.76s/it]INFO 04-12 13:23:07 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 657/3300 [2:37:06<11:37:12, 15.83s/it]2026-04-12 13:23:18,242 | INFO | train_grpo_train | metrics_logged mode=train step=657

                                                       
{'loss': -0.1019, 'grad_norm': 8.138545989990234, 'learning_rate': 8.012121212121214e-06, 'num_tokens': 1318125.0, 'completions/mean_length': 127.25, 'completions/min_length': 42.0, 'completions/max_length': 155.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 127.25, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 155.0, 'rewards/meter/mean': 0.43511897325515747, 'rewards/meter/std': 0.2925766110420227, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.2182178944349289, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9402433037757874, 'rewards/lexical_plausibility/std': 0.13108384609222412, 'rewards/judge_quality/mean': 0.23750001192092896, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.6454905271530151, 'rewards/repeat_penalty/std': 0.28969377279281616, 'rewards/near_duplicate_penalty/mean': 0.8732624053955078, 'rewards/near_duplicate_penalty/std': 0.2218487709760666, 'rewards/opening_diversity/mean': 0.846875011920929, 'rewards/opening_diversity/std': 0.3048528730869293, 'rewards/distinct_2/mean': 0.7203527092933655, 'rewards/distinct_2/std': 0.30008402466773987, 'rewards/total_composite/mean': 0.08786556124687195, 'rewards/total_composite/std': 0.062393952161073685, 'reward': 0.08786556124687195, 'reward_std': 0.062393948435783386, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14421558380126953, 'sampling/sampling_logp_difference/max': 2.2183570861816406, 'sampling/importance_sampling_ratio/min': 0.10878768563270569, 'sampling/importance_sampling_ratio/mean': 0.9979930520057678, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8150635771453381, 'clip_ratio/low_mean': 0.04829929443076253, 'clip_ratio/low_min': 0.04829929443076253, 'clip_ratio/high_mean': 0.07662495132535696, 'clip_ratio/high_max': 0.07662495132535696, 'clip_ratio/region_mean': 0.12492424575611949, 'reward_total_mean': 0.08786556124687195, 'reward_meter_mean': 0.43511897325515747, 'reward_meter_std': 0.2925766110420227, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.2182178944349289, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9402433037757874, 'reward_lexical_plausibility_std': 0.13108384609222412, 'reward_repeat_penalty_mean': 0.6454905271530151, 'reward_repeat_penalty_std': 0.28969377279281616, 'reward_near_duplicate_penalty_mean': 0.8732624053955078, 'reward_near_duplicate_penalty_std': 0.2218487709760666, 'reward_opening_diversity_mean': 0.846875011920929, 'reward_opening_diversity_std': 0.3048528730869293, 'reward_distinct_2_mean': 0.7203527092933655, 'reward_distinct_2_std': 0.30008402466773987, 'reward_judge_quality_mean': 0.23750001192092896, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.08786556124687195, 'reward_total_composite_std': 0.062393952161073685, 'epoch': 0.03}

 20%|█▉        | 657/3300 [2:37:06<11:37:12, 15.83s/it]INFO 04-12 13:23:18 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 658/3300 [2:37:21<11:27:02, 15.60s/it]2026-04-12 13:23:33,319 | INFO | train_grpo_train | metrics_logged mode=train step=658

                                                       
{'loss': 0.0301, 'grad_norm': 2.0185253620147705, 'learning_rate': 8.00909090909091e-06, 'num_tokens': 1323693.0, 'completions/mean_length': 451.0, 'completions/min_length': 407.0, 'completions/max_length': 489.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 451.0, 'completions/min_terminated_length': 407.0, 'completions/max_terminated_length': 489.0, 'rewards/meter/mean': 0.5132176876068115, 'rewards/meter/std': 0.483419269323349, 'rewards/count_adherence/mean': 0.7916666269302368, 'rewards/count_adherence/std': 0.04272466152906418, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.125, 'rewards/judge_quality/std': 0.046291008591651917, 'rewards/repeat_penalty/mean': 0.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 0.4448012411594391, 'rewards/near_duplicate_penalty/std': 0.2239237129688263, 'rewards/opening_diversity/mean': 0.7188467979431152, 'rewards/opening_diversity/std': 0.2755339741706848, 'rewards/distinct_2/mean': 0.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.05027870833873749, 'rewards/total_composite/std': 0.0547531358897686, 'reward': 0.05027870833873749, 'reward_std': 0.0547531358897686, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02426569163799286, 'sampling/sampling_logp_difference/max': 2.7454569339752197, 'sampling/importance_sampling_ratio/min': 0.06421895325183868, 'sampling/importance_sampling_ratio/mean': 1.0036330223083496, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1453683408908546, 'clip_ratio/low_mean': 0.013538115366827697, 'clip_ratio/low_min': 0.013538115366827697, 'clip_ratio/high_mean': 0.01030551188159734, 'clip_ratio/high_max': 0.01030551188159734, 'clip_ratio/region_mean': 0.023843627248425037, 'reward_total_mean': 0.05027870833873749, 'reward_meter_mean': 0.5132176876068115, 'reward_meter_std': 0.483419269323349, 'reward_count_adherence_mean': 0.7916666269302368, 'reward_count_adherence_std': 0.04272466152906418, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.0, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 0.4448012411594391, 'reward_near_duplicate_penalty_std': 0.2239237129688263, 'reward_opening_diversity_mean': 0.7188467979431152, 'reward_opening_diversity_std': 0.2755339741706848, 'reward_distinct_2_mean': 0.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.125, 'reward_judge_quality_std': 0.046291008591651917, 'reward_total_composite_mean': 0.05027870833873749, 'reward_total_composite_std': 0.0547531358897686, 'epoch': 0.03}

 20%|█▉        | 658/3300 [2:37:21<11:27:02, 15.60s/it]INFO 04-12 13:23:33 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 659/3300 [2:37:35<11:12:53, 15.29s/it]2026-04-12 13:23:47,871 | INFO | train_grpo_train | metrics_logged mode=train step=659

                                                       
{'loss': -0.1073, 'grad_norm': 5.058903217315674, 'learning_rate': 8.006060606060607e-06, 'num_tokens': 1325330.0, 'completions/mean_length': 111.625, 'completions/min_length': 43.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 54.42857360839844, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.7167664766311646, 'rewards/meter/std': 0.40534457564353943, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9550361633300781, 'rewards/lexical_plausibility/std': 0.12717685103416443, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.8984770178794861, 'rewards/repeat_penalty/std': 0.13434016704559326, 'rewards/near_duplicate_penalty/mean': 0.9604055881500244, 'rewards/near_duplicate_penalty/std': 0.03500939533114433, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9608974456787109, 'rewards/distinct_2/std': 0.057494573295116425, 'rewards/total_composite/mean': 0.24018089473247528, 'rewards/total_composite/std': 0.15819303691387177, 'reward': 0.24018089473247528, 'reward_std': 0.15819303691387177, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16321519017219543, 'sampling/sampling_logp_difference/max': 1.762455940246582, 'sampling/importance_sampling_ratio/min': 0.1716228574514389, 'sampling/importance_sampling_ratio/mean': 1.0060358047485352, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7552897930145264, 'clip_ratio/low_mean': 0.01922682556323707, 'clip_ratio/low_min': 0.01922682556323707, 'clip_ratio/high_mean': 0.0866461917757988, 'clip_ratio/high_max': 0.0866461917757988, 'clip_ratio/region_mean': 0.10587301733903587, 'reward_total_mean': 0.24018089473247528, 'reward_meter_mean': 0.7167664766311646, 'reward_meter_std': 0.40534457564353943, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9550361633300781, 'reward_lexical_plausibility_std': 0.12717685103416443, 'reward_repeat_penalty_mean': 0.8984770178794861, 'reward_repeat_penalty_std': 0.13434016704559326, 'reward_near_duplicate_penalty_mean': 0.9604055881500244, 'reward_near_duplicate_penalty_std': 0.03500939533114433, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9608974456787109, 'reward_distinct_2_std': 0.057494573295116425, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.24018089473247528, 'reward_total_composite_std': 0.15819303691387177, 'epoch': 0.03}

 20%|█▉        | 659/3300 [2:37:35<11:12:53, 15.29s/it]INFO 04-12 13:23:48 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 660/3300 [2:37:45<9:54:56, 13.52s/it] 2026-04-12 13:23:57,273 | INFO | train_grpo_train | metrics_logged mode=train step=660

                                                      
{'loss': 0.0499, 'grad_norm': 10.33723258972168, 'learning_rate': 8.003030303030304e-06, 'num_tokens': 1327271.0, 'completions/mean_length': 85.625, 'completions/min_length': 78.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 85.625, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.6358233690261841, 'rewards/meter/std': 0.34739404916763306, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9977678656578064, 'rewards/lexical_plausibility/std': 0.006313450634479523, 'rewards/judge_quality/mean': 0.5449999570846558, 'rewards/judge_quality/std': 0.22915996611118317, 'rewards/repeat_penalty/mean': 0.9449267387390137, 'rewards/repeat_penalty/std': 0.03723674267530441, 'rewards/near_duplicate_penalty/mean': 0.9611337184906006, 'rewards/near_duplicate_penalty/std': 0.016394222155213356, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9828397631645203, 'rewards/distinct_2/std': 0.02371014468371868, 'rewards/total_composite/mean': 0.327256977558136, 'rewards/total_composite/std': 0.23568442463874817, 'reward': 0.327256977558136, 'reward_std': 0.23568440973758698, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15584082901477814, 'sampling/sampling_logp_difference/max': 3.2166783809661865, 'sampling/importance_sampling_ratio/min': 0.0400879941880703, 'sampling/importance_sampling_ratio/mean': 1.0125001668930054, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8571107685565948, 'clip_ratio/low_mean': 0.07578548043966293, 'clip_ratio/low_min': 0.07578548043966293, 'clip_ratio/high_mean': 0.05673158261924982, 'clip_ratio/high_max': 0.05673158261924982, 'clip_ratio/region_mean': 0.13251706305891275, 'reward_total_mean': 0.327256977558136, 'reward_meter_mean': 0.6358233690261841, 'reward_meter_std': 0.34739404916763306, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9977678656578064, 'reward_lexical_plausibility_std': 0.006313450634479523, 'reward_repeat_penalty_mean': 0.9449267387390137, 'reward_repeat_penalty_std': 0.03723674267530441, 'reward_near_duplicate_penalty_mean': 0.9611337184906006, 'reward_near_duplicate_penalty_std': 0.016394222155213356, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9828397631645203, 'reward_distinct_2_std': 0.02371014468371868, 'reward_judge_quality_mean': 0.5449999570846558, 'reward_judge_quality_std': 0.22915996611118317, 'reward_total_composite_mean': 0.327256977558136, 'reward_total_composite_std': 0.23568442463874817, 'epoch': 0.03}

 20%|██        | 660/3300 [2:37:45<9:54:56, 13.52s/it]INFO 04-12 13:23:57 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 661/3300 [2:37:55<9:14:28, 12.61s/it]2026-04-12 13:24:07,744 | INFO | train_grpo_train | metrics_logged mode=train step=661

                                                      
{'loss': -0.0203, 'grad_norm': 11.228118896484375, 'learning_rate': 8.000000000000001e-06, 'num_tokens': 1329368.0, 'completions/mean_length': 79.125, 'completions/min_length': 45.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.125, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.35494035482406616, 'rewards/meter/std': 0.38554656505584717, '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/lexical_plausibility/mean': 0.9864042401313782, 'rewards/lexical_plausibility/std': 0.031858112663030624, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.08345229178667068, 'rewards/repeat_penalty/mean': 0.9130034446716309, 'rewards/repeat_penalty/std': 0.1118207573890686, 'rewards/near_duplicate_penalty/mean': 0.9660369753837585, 'rewards/near_duplicate_penalty/std': 0.024087518453598022, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9431574940681458, 'rewards/distinct_2/std': 0.0974927470088005, 'rewards/total_composite/mean': 0.11021161079406738, 'rewards/total_composite/std': 0.13464927673339844, 'reward': 0.11021161079406738, 'reward_std': 0.13464926183223724, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15921148657798767, 'sampling/sampling_logp_difference/max': 2.1211700439453125, 'sampling/importance_sampling_ratio/min': 0.11989127099514008, 'sampling/importance_sampling_ratio/mean': 0.9796518683433533, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.687077060341835, 'clip_ratio/low_mean': 0.09710708633065224, 'clip_ratio/low_min': 0.09710708633065224, 'clip_ratio/high_mean': 0.031055900268256664, 'clip_ratio/high_max': 0.031055900268256664, 'clip_ratio/region_mean': 0.1281629865989089, 'reward_total_mean': 0.11021161079406738, 'reward_meter_mean': 0.35494035482406616, 'reward_meter_std': 0.38554656505584717, '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_lexical_plausibility_mean': 0.9864042401313782, 'reward_lexical_plausibility_std': 0.031858112663030624, 'reward_repeat_penalty_mean': 0.9130034446716309, 'reward_repeat_penalty_std': 0.1118207573890686, 'reward_near_duplicate_penalty_mean': 0.9660369753837585, 'reward_near_duplicate_penalty_std': 0.024087518453598022, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9431574940681458, 'reward_distinct_2_std': 0.0974927470088005, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.08345229178667068, 'reward_total_composite_mean': 0.11021161079406738, 'reward_total_composite_std': 0.13464927673339844, 'epoch': 0.03}

 20%|██        | 661/3300 [2:37:55<9:14:28, 12.61s/it]INFO 04-12 13:24:07 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 662/3300 [2:38:09<9:32:44, 13.03s/it]2026-04-12 13:24:21,751 | INFO | train_grpo_train | metrics_logged mode=train step=662

                                                      
{'loss': -0.0499, 'grad_norm': 5.116161346435547, 'learning_rate': 7.996969696969697e-06, 'num_tokens': 1331139.0, 'completions/mean_length': 102.375, 'completions/min_length': 38.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 43.85714340209961, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.4327571988105774, 'rewards/meter/std': 0.4184527099132538, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9693456888198853, 'rewards/lexical_plausibility/std': 0.08670346438884735, 'rewards/judge_quality/mean': 0.6937500238418579, 'rewards/judge_quality/std': 0.335769385099411, 'rewards/repeat_penalty/mean': 0.8699858784675598, 'rewards/repeat_penalty/std': 0.1817188262939453, 'rewards/near_duplicate_penalty/mean': 0.9534045457839966, 'rewards/near_duplicate_penalty/std': 0.062445808202028275, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9624541997909546, 'rewards/distinct_2/std': 0.0560930110514164, 'rewards/total_composite/mean': 0.3592093586921692, 'rewards/total_composite/std': 0.4135020971298218, 'reward': 0.3592093586921692, 'reward_std': 0.4135020971298218, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13316769897937775, 'sampling/sampling_logp_difference/max': 1.3598146438598633, 'sampling/importance_sampling_ratio/min': 0.2567083537578583, 'sampling/importance_sampling_ratio/mean': 1.0052872896194458, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.690619707107544, 'clip_ratio/low_mean': 0.052816879004240036, 'clip_ratio/low_min': 0.052816879004240036, 'clip_ratio/high_mean': 0.0431758863851428, 'clip_ratio/high_max': 0.0431758863851428, 'clip_ratio/region_mean': 0.09599276538938284, 'reward_total_mean': 0.3592093586921692, 'reward_meter_mean': 0.4327571988105774, 'reward_meter_std': 0.4184527099132538, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9693456888198853, 'reward_lexical_plausibility_std': 0.08670346438884735, 'reward_repeat_penalty_mean': 0.8699858784675598, 'reward_repeat_penalty_std': 0.1817188262939453, 'reward_near_duplicate_penalty_mean': 0.9534045457839966, 'reward_near_duplicate_penalty_std': 0.062445808202028275, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9624541997909546, 'reward_distinct_2_std': 0.0560930110514164, 'reward_judge_quality_mean': 0.6937500238418579, 'reward_judge_quality_std': 0.335769385099411, 'reward_total_composite_mean': 0.3592093586921692, 'reward_total_composite_std': 0.4135020971298218, 'epoch': 0.03}

 20%|██        | 662/3300 [2:38:09<9:32:44, 13.03s/it]INFO 04-12 13:24:21 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 663/3300 [2:38:19<8:53:56, 12.15s/it]2026-04-12 13:24:31,850 | INFO | train_grpo_train | metrics_logged mode=train step=663

                                                      
{'loss': 0.062, 'grad_norm': 8.638606071472168, 'learning_rate': 7.993939393939396e-06, 'num_tokens': 1333388.0, 'completions/mean_length': 100.125, 'completions/min_length': 91.0, 'completions/max_length': 110.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 100.125, 'completions/min_terminated_length': 91.0, 'completions/max_terminated_length': 110.0, 'rewards/meter/mean': 0.6953133344650269, 'rewards/meter/std': 0.36091312766075134, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.1414213478565216, 'rewards/repeat_penalty/mean': 0.8255414366722107, 'rewards/repeat_penalty/std': 0.11612191051244736, 'rewards/near_duplicate_penalty/mean': 0.9499050378799438, 'rewards/near_duplicate_penalty/std': 0.03630257397890091, 'rewards/opening_diversity/mean': 0.984375, 'rewards/opening_diversity/std': 0.0289318785071373, 'rewards/distinct_2/mean': 0.8796696662902832, 'rewards/distinct_2/std': 0.08587934076786041, 'rewards/total_composite/mean': 0.2767334580421448, 'rewards/total_composite/std': 0.1958579123020172, 'reward': 0.2767334580421448, 'reward_std': 0.1958579123020172, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1456546187400818, 'sampling/sampling_logp_difference/max': 1.9623117446899414, 'sampling/importance_sampling_ratio/min': 0.14053316414356232, 'sampling/importance_sampling_ratio/mean': 1.00351083278656, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6855262108147144, 'clip_ratio/low_mean': 0.06558650452643633, 'clip_ratio/low_min': 0.06558650452643633, 'clip_ratio/high_mean': 0.06632081046700478, 'clip_ratio/high_max': 0.06632081046700478, 'clip_ratio/region_mean': 0.1319073149934411, 'reward_total_mean': 0.2767334580421448, 'reward_meter_mean': 0.6953133344650269, 'reward_meter_std': 0.36091312766075134, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8255414366722107, 'reward_repeat_penalty_std': 0.11612191051244736, 'reward_near_duplicate_penalty_mean': 0.9499050378799438, 'reward_near_duplicate_penalty_std': 0.03630257397890091, 'reward_opening_diversity_mean': 0.984375, 'reward_opening_diversity_std': 0.0289318785071373, 'reward_distinct_2_mean': 0.8796696662902832, 'reward_distinct_2_std': 0.08587934076786041, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.1414213478565216, 'reward_total_composite_mean': 0.2767334580421448, 'reward_total_composite_std': 0.1958579123020172, 'epoch': 0.03}

 20%|██        | 663/3300 [2:38:19<8:53:56, 12.15s/it]INFO 04-12 13:24:32 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 664/3300 [2:38:28<8:13:59, 11.24s/it]2026-04-12 13:24:40,982 | INFO | train_grpo_train | metrics_logged mode=train step=664

                                                      
{'loss': -0.0089, 'grad_norm': 13.82679271697998, 'learning_rate': 7.990909090909091e-06, 'num_tokens': 1335064.0, 'completions/mean_length': 53.5, 'completions/min_length': 48.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.5, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.7547950744628906, 'rewards/meter/std': 0.29388749599456787, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6124999523162842, 'rewards/judge_quality/std': 0.23741163313388824, 'rewards/repeat_penalty/mean': 0.914587140083313, 'rewards/repeat_penalty/std': 0.13662421703338623, 'rewards/near_duplicate_penalty/mean': 0.9572491645812988, 'rewards/near_duplicate_penalty/std': 0.018889212980866432, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9797570705413818, 'rewards/distinct_2/std': 0.05725561082363129, 'rewards/total_composite/mean': 0.4655528962612152, 'rewards/total_composite/std': 0.2876104414463043, 'reward': 0.4655528962612152, 'reward_std': 0.28761041164398193, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1815180480480194, 'sampling/sampling_logp_difference/max': 5.202572822570801, 'sampling/importance_sampling_ratio/min': 0.005502389278262854, 'sampling/importance_sampling_ratio/mean': 1.001611590385437, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8862349018454552, 'clip_ratio/low_mean': 0.06500632595270872, 'clip_ratio/low_min': 0.06500632595270872, 'clip_ratio/high_mean': 0.09317438304424286, 'clip_ratio/high_max': 0.09317438304424286, 'clip_ratio/region_mean': 0.15818070899695158, 'reward_total_mean': 0.4655528962612152, 'reward_meter_mean': 0.7547950744628906, 'reward_meter_std': 0.29388749599456787, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.914587140083313, 'reward_repeat_penalty_std': 0.13662421703338623, 'reward_near_duplicate_penalty_mean': 0.9572491645812988, 'reward_near_duplicate_penalty_std': 0.018889212980866432, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9797570705413818, 'reward_distinct_2_std': 0.05725561082363129, 'reward_judge_quality_mean': 0.6124999523162842, 'reward_judge_quality_std': 0.23741163313388824, 'reward_total_composite_mean': 0.4655528962612152, 'reward_total_composite_std': 0.2876104414463043, 'epoch': 0.03}

 20%|██        | 664/3300 [2:38:28<8:13:59, 11.24s/it]INFO 04-12 13:24:41 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 665/3300 [2:38:37<7:41:36, 10.51s/it]2026-04-12 13:24:49,785 | INFO | train_grpo_train | metrics_logged mode=train step=665

                                                      
{'loss': 0.0249, 'grad_norm': 13.572747230529785, 'learning_rate': 7.987878787878789e-06, 'num_tokens': 1336727.0, 'completions/mean_length': 56.875, 'completions/min_length': 47.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.875, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.8040631413459778, 'rewards/meter/std': 0.33520224690437317, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.8539631366729736, 'rewards/lexical_plausibility/std': 0.1430853307247162, 'rewards/judge_quality/mean': 0.5350000262260437, 'rewards/judge_quality/std': 0.411582350730896, 'rewards/repeat_penalty/mean': 0.964275598526001, 'rewards/repeat_penalty/std': 0.05441336706280708, 'rewards/near_duplicate_penalty/mean': 0.9808967113494873, 'rewards/near_duplicate_penalty/std': 0.03064265102148056, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9825174808502197, 'rewards/distinct_2/std': 0.03237133473157883, 'rewards/total_composite/mean': 0.4284802973270416, 'rewards/total_composite/std': 0.4009256958961487, 'reward': 0.4284802973270416, 'reward_std': 0.4009256958961487, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1456071138381958, 'sampling/sampling_logp_difference/max': 1.9290480613708496, 'sampling/importance_sampling_ratio/min': 0.1452864408493042, 'sampling/importance_sampling_ratio/mean': 1.0067170858383179, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9042868167161942, 'clip_ratio/low_mean': 0.08205477707087994, 'clip_ratio/low_min': 0.08205477707087994, 'clip_ratio/high_mean': 0.05415690876543522, 'clip_ratio/high_max': 0.05415690876543522, 'clip_ratio/region_mean': 0.13621168583631516, 'reward_total_mean': 0.4284802973270416, 'reward_meter_mean': 0.8040631413459778, 'reward_meter_std': 0.33520224690437317, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.8539631366729736, 'reward_lexical_plausibility_std': 0.1430853307247162, 'reward_repeat_penalty_mean': 0.964275598526001, 'reward_repeat_penalty_std': 0.05441336706280708, 'reward_near_duplicate_penalty_mean': 0.9808967113494873, 'reward_near_duplicate_penalty_std': 0.03064265102148056, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9825174808502197, 'reward_distinct_2_std': 0.03237133473157883, 'reward_judge_quality_mean': 0.5350000262260437, 'reward_judge_quality_std': 0.411582350730896, 'reward_total_composite_mean': 0.4284802973270416, 'reward_total_composite_std': 0.4009256958961487, 'epoch': 0.03}

 20%|██        | 665/3300 [2:38:37<7:41:36, 10.51s/it]INFO 04-12 13:24:50 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 666/3300 [2:38:47<7:34:37, 10.36s/it]2026-04-12 13:24:59,781 | INFO | train_grpo_train | metrics_logged mode=train step=666

                                                      
{'loss': 0.0983, 'grad_norm': 16.244979858398438, 'learning_rate': 7.984848484848486e-06, 'num_tokens': 1338812.0, 'completions/mean_length': 93.625, 'completions/min_length': 72.0, 'completions/max_length': 103.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 93.625, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.6764025688171387, 'rewards/meter/std': 0.27123498916625977, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4624999761581421, 'rewards/judge_quality/std': 0.15526476502418518, 'rewards/repeat_penalty/mean': 0.9311668872833252, 'rewards/repeat_penalty/std': 0.06370653957128525, 'rewards/near_duplicate_penalty/mean': 0.9684038162231445, 'rewards/near_duplicate_penalty/std': 0.021242352202534676, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9611342549324036, 'rewards/distinct_2/std': 0.05580252408981323, 'rewards/total_composite/mean': 0.2956639230251312, 'rewards/total_composite/std': 0.12961456179618835, 'reward': 0.2956639230251312, 'reward_std': 0.12961454689502716, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14634813368320465, 'sampling/sampling_logp_difference/max': 1.6656081676483154, 'sampling/importance_sampling_ratio/min': 0.18907563388347626, 'sampling/importance_sampling_ratio/mean': 1.015251874923706, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7250681594014168, 'clip_ratio/low_mean': 0.061393896117806435, 'clip_ratio/low_min': 0.061393896117806435, 'clip_ratio/high_mean': 0.08866982720792294, 'clip_ratio/high_max': 0.08866982720792294, 'clip_ratio/region_mean': 0.15006372332572937, 'reward_total_mean': 0.2956639230251312, 'reward_meter_mean': 0.6764025688171387, 'reward_meter_std': 0.27123498916625977, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9311668872833252, 'reward_repeat_penalty_std': 0.06370653957128525, 'reward_near_duplicate_penalty_mean': 0.9684038162231445, 'reward_near_duplicate_penalty_std': 0.021242352202534676, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9611342549324036, 'reward_distinct_2_std': 0.05580252408981323, 'reward_judge_quality_mean': 0.4624999761581421, 'reward_judge_quality_std': 0.15526476502418518, 'reward_total_composite_mean': 0.2956639230251312, 'reward_total_composite_std': 0.12961456179618835, 'epoch': 0.03}

 20%|██        | 666/3300 [2:38:47<7:34:37, 10.36s/it]INFO 04-12 13:25:00 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 667/3300 [2:38:58<7:42:50, 10.55s/it]2026-04-12 13:25:10,773 | INFO | train_grpo_train | metrics_logged mode=train step=667

                                                      
{'loss': 0.0539, 'grad_norm': 11.671969413757324, 'learning_rate': 7.981818181818183e-06, 'num_tokens': 1340980.0, 'completions/mean_length': 94.0, 'completions/min_length': 84.0, 'completions/max_length': 108.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 94.0, 'completions/min_terminated_length': 84.0, 'completions/max_terminated_length': 108.0, 'rewards/meter/mean': 0.5911504030227661, 'rewards/meter/std': 0.3050113022327423, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.9143061637878418, 'rewards/repeat_penalty/std': 0.056243348866701126, 'rewards/near_duplicate_penalty/mean': 0.9560746550559998, 'rewards/near_duplicate_penalty/std': 0.03006063774228096, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9558618068695068, 'rewards/distinct_2/std': 0.03913449868559837, 'rewards/total_composite/mean': 0.15422356128692627, 'rewards/total_composite/std': 0.1303739994764328, 'reward': 0.15422356128692627, 'reward_std': 0.1303739994764328, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1310657262802124, 'sampling/sampling_logp_difference/max': 2.116398811340332, 'sampling/importance_sampling_ratio/min': 0.12046466022729874, 'sampling/importance_sampling_ratio/mean': 1.012769341468811, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.626823864877224, 'clip_ratio/low_mean': 0.07949473522603512, 'clip_ratio/low_min': 0.07949473522603512, 'clip_ratio/high_mean': 0.042927565053105354, 'clip_ratio/high_max': 0.042927565053105354, 'clip_ratio/region_mean': 0.12242230027914047, 'reward_total_mean': 0.15422356128692627, 'reward_meter_mean': 0.5911504030227661, 'reward_meter_std': 0.3050113022327423, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9143061637878418, 'reward_repeat_penalty_std': 0.056243348866701126, 'reward_near_duplicate_penalty_mean': 0.9560746550559998, 'reward_near_duplicate_penalty_std': 0.03006063774228096, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9558618068695068, 'reward_distinct_2_std': 0.03913449868559837, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.15422356128692627, 'reward_total_composite_std': 0.1303739994764328, 'epoch': 0.03}

 20%|██        | 667/3300 [2:38:58<7:42:50, 10.55s/it]INFO 04-12 13:25:11 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 668/3300 [2:39:09<7:40:59, 10.51s/it]2026-04-12 13:25:21,193 | INFO | train_grpo_train | metrics_logged mode=train step=668

                                                      
{'loss': -0.0223, 'grad_norm': 8.770500183105469, 'learning_rate': 7.978787878787879e-06, 'num_tokens': 1343552.0, 'completions/mean_length': 123.5, 'completions/min_length': 96.0, 'completions/max_length': 132.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 123.5, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 132.0, 'rewards/meter/mean': 0.36670535802841187, 'rewards/meter/std': 0.363443523645401, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.17500001192092896, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/repeat_penalty/mean': 0.6017524003982544, 'rewards/repeat_penalty/std': 0.17603251338005066, 'rewards/near_duplicate_penalty/mean': 0.8795514702796936, 'rewards/near_duplicate_penalty/std': 0.04831100255250931, 'rewards/opening_diversity/mean': 0.987500011920929, 'rewards/opening_diversity/std': 0.0353553481400013, 'rewards/distinct_2/mean': 0.6907018423080444, 'rewards/distinct_2/std': 0.14063426852226257, 'rewards/total_composite/mean': 0.06219920516014099, 'rewards/total_composite/std': 0.06289541721343994, 'reward': 0.06219920516014099, 'reward_std': 0.06289542466402054, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1278284341096878, 'sampling/sampling_logp_difference/max': 2.4323699474334717, 'sampling/importance_sampling_ratio/min': 0.08782843500375748, 'sampling/importance_sampling_ratio/mean': 0.9909975528717041, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6201633624732494, 'clip_ratio/low_mean': 0.06598906312137842, 'clip_ratio/low_min': 0.06598906312137842, 'clip_ratio/high_mean': 0.04261363670229912, 'clip_ratio/high_max': 0.04261363670229912, 'clip_ratio/region_mean': 0.10860269982367754, 'reward_total_mean': 0.06219920516014099, 'reward_meter_mean': 0.36670535802841187, 'reward_meter_std': 0.363443523645401, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.0707106739282608, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6017524003982544, 'reward_repeat_penalty_std': 0.17603251338005066, 'reward_near_duplicate_penalty_mean': 0.8795514702796936, 'reward_near_duplicate_penalty_std': 0.04831100255250931, 'reward_opening_diversity_mean': 0.987500011920929, 'reward_opening_diversity_std': 0.0353553481400013, 'reward_distinct_2_mean': 0.6907018423080444, 'reward_distinct_2_std': 0.14063426852226257, 'reward_judge_quality_mean': 0.17500001192092896, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.06219920516014099, 'reward_total_composite_std': 0.06289541721343994, 'epoch': 0.03}

 20%|██        | 668/3300 [2:39:09<7:40:59, 10.51s/it]INFO 04-12 13:25:21 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 669/3300 [2:39:19<7:37:30, 10.43s/it]2026-04-12 13:25:31,449 | INFO | train_grpo_train | metrics_logged mode=train step=669

                                                      
{'loss': 0.0056, 'grad_norm': 13.67820930480957, 'learning_rate': 7.975757575757576e-06, 'num_tokens': 1345383.0, 'completions/mean_length': 51.875, 'completions/min_length': 45.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.875, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.7102116346359253, 'rewards/meter/std': 0.38947975635528564, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.48374998569488525, 'rewards/judge_quality/std': 0.18196842074394226, 'rewards/repeat_penalty/mean': 0.9329346418380737, 'rewards/repeat_penalty/std': 0.05199816823005676, 'rewards/near_duplicate_penalty/mean': 0.95139479637146, 'rewards/near_duplicate_penalty/std': 0.030207542702555656, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9802631139755249, 'rewards/distinct_2/std': 0.036561544984579086, 'rewards/total_composite/mean': 0.32223033905029297, 'rewards/total_composite/std': 0.15651214122772217, 'reward': 0.32223033905029297, 'reward_std': 0.15651212632656097, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14197148382663727, 'sampling/sampling_logp_difference/max': 2.001180648803711, 'sampling/importance_sampling_ratio/min': 0.13517560064792633, 'sampling/importance_sampling_ratio/mean': 0.9927393198013306, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7489805072546005, 'clip_ratio/low_mean': 0.036504664458334446, 'clip_ratio/low_min': 0.036504664458334446, 'clip_ratio/high_mean': 0.08419650979340076, 'clip_ratio/high_max': 0.08419650979340076, 'clip_ratio/region_mean': 0.12070117425173521, 'reward_total_mean': 0.32223033905029297, 'reward_meter_mean': 0.7102116346359253, 'reward_meter_std': 0.38947975635528564, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9329346418380737, 'reward_repeat_penalty_std': 0.05199816823005676, 'reward_near_duplicate_penalty_mean': 0.95139479637146, 'reward_near_duplicate_penalty_std': 0.030207542702555656, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9802631139755249, 'reward_distinct_2_std': 0.036561544984579086, 'reward_judge_quality_mean': 0.48374998569488525, 'reward_judge_quality_std': 0.18196842074394226, 'reward_total_composite_mean': 0.32223033905029297, 'reward_total_composite_std': 0.15651214122772217, 'epoch': 0.03}

 20%|██        | 669/3300 [2:39:19<7:37:30, 10.43s/it]INFO 04-12 13:25:31 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 670/3300 [2:39:28<7:22:00, 10.08s/it]2026-04-12 13:25:40,718 | INFO | train_grpo_train | metrics_logged mode=train step=670

                                                      
{'loss': 0.0172, 'grad_norm': 12.20312786102295, 'learning_rate': 7.972727272727273e-06, 'num_tokens': 1347125.0, 'completions/mean_length': 52.75, 'completions/min_length': 48.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.75, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.987130880355835, 'rewards/meter/std': 0.0074296449311077595, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9887152910232544, 'rewards/lexical_plausibility/std': 0.022863244637846947, 'rewards/judge_quality/mean': 0.6012499928474426, 'rewards/judge_quality/std': 0.2671777307987213, 'rewards/repeat_penalty/mean': 0.8460367918014526, 'rewards/repeat_penalty/std': 0.1868453323841095, 'rewards/near_duplicate_penalty/mean': 0.9336796998977661, 'rewards/near_duplicate_penalty/std': 0.07384230196475983, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9526585340499878, 'rewards/distinct_2/std': 0.055468734353780746, 'rewards/total_composite/mean': 0.5933990478515625, 'rewards/total_composite/std': 0.2635002136230469, 'reward': 0.5933990478515625, 'reward_std': 0.2635002136230469, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1362331360578537, 'sampling/sampling_logp_difference/max': 1.7461051940917969, 'sampling/importance_sampling_ratio/min': 0.1798045039176941, 'sampling/importance_sampling_ratio/mean': 1.0128592252731323, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7716864421963692, 'clip_ratio/low_mean': 0.053620395716279745, 'clip_ratio/low_min': 0.053620395716279745, 'clip_ratio/high_mean': 0.03396551730111241, 'clip_ratio/high_max': 0.03396551730111241, 'clip_ratio/region_mean': 0.08758591301739216, 'reward_total_mean': 0.5933990478515625, 'reward_meter_mean': 0.987130880355835, 'reward_meter_std': 0.0074296449311077595, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9887152910232544, 'reward_lexical_plausibility_std': 0.022863244637846947, 'reward_repeat_penalty_mean': 0.8460367918014526, 'reward_repeat_penalty_std': 0.1868453323841095, 'reward_near_duplicate_penalty_mean': 0.9336796998977661, 'reward_near_duplicate_penalty_std': 0.07384230196475983, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9526585340499878, 'reward_distinct_2_std': 0.055468734353780746, 'reward_judge_quality_mean': 0.6012499928474426, 'reward_judge_quality_std': 0.2671777307987213, 'reward_total_composite_mean': 0.5933990478515625, 'reward_total_composite_std': 0.2635002136230469, 'epoch': 0.03}

 20%|██        | 670/3300 [2:39:28<7:22:00, 10.08s/it]INFO 04-12 13:25:40 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 671/3300 [2:39:38<7:24:32, 10.15s/it]2026-04-12 13:25:51,007 | INFO | train_grpo_train | metrics_logged mode=train step=671

                                                      
{'loss': 0.0711, 'grad_norm': 9.78178882598877, 'learning_rate': 7.96969696969697e-06, 'num_tokens': 1349057.0, 'completions/mean_length': 72.5, 'completions/min_length': 63.0, 'completions/max_length': 84.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.5, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.3886287808418274, 'rewards/meter/std': 0.40985411405563354, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.05345224589109421, 'rewards/repeat_penalty/mean': 0.9256945848464966, 'rewards/repeat_penalty/std': 0.04860997945070267, 'rewards/near_duplicate_penalty/mean': 0.9604939818382263, 'rewards/near_duplicate_penalty/std': 0.02726934663951397, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.963475227355957, 'rewards/distinct_2/std': 0.03487830236554146, 'rewards/total_composite/mean': 0.1587459146976471, 'rewards/total_composite/std': 0.17057305574417114, 'reward': 0.1587459146976471, 'reward_std': 0.17057305574417114, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1335901916027069, 'sampling/sampling_logp_difference/max': 2.0363283157348633, 'sampling/importance_sampling_ratio/min': 0.13050702214241028, 'sampling/importance_sampling_ratio/mean': 1.0084887742996216, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7365903630852699, 'clip_ratio/low_mean': 0.0709469704888761, 'clip_ratio/low_min': 0.0709469704888761, 'clip_ratio/high_mean': 0.042468528263270855, 'clip_ratio/high_max': 0.042468528263270855, 'clip_ratio/region_mean': 0.11341549875214696, 'reward_total_mean': 0.1587459146976471, 'reward_meter_mean': 0.3886287808418274, 'reward_meter_std': 0.40985411405563354, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9256945848464966, 'reward_repeat_penalty_std': 0.04860997945070267, 'reward_near_duplicate_penalty_mean': 0.9604939818382263, 'reward_near_duplicate_penalty_std': 0.02726934663951397, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.963475227355957, 'reward_distinct_2_std': 0.03487830236554146, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.05345224589109421, 'reward_total_composite_mean': 0.1587459146976471, 'reward_total_composite_std': 0.17057305574417114, 'epoch': 0.03}

 20%|██        | 671/3300 [2:39:38<7:24:32, 10.15s/it]INFO 04-12 13:25:51 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 672/3300 [2:39:54<8:36:12, 11.79s/it]2026-04-12 13:26:06,621 | INFO | train_grpo_train | metrics_logged mode=train step=672

                                                      
{'loss': -0.0295, 'grad_norm': 7.1071672439575195, 'learning_rate': 7.966666666666668e-06, 'num_tokens': 1351271.0, 'completions/mean_length': 162.75, 'completions/min_length': 85.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 112.85714721679688, 'completions/min_terminated_length': 85.0, 'completions/max_terminated_length': 141.0, 'rewards/meter/mean': 0.4819783866405487, 'rewards/meter/std': 0.3900616765022278, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/repeat_penalty/mean': 0.7841250896453857, 'rewards/repeat_penalty/std': 0.3230384290218353, 'rewards/near_duplicate_penalty/mean': 0.9610974192619324, 'rewards/near_duplicate_penalty/std': 0.01954832673072815, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.816798746585846, 'rewards/distinct_2/std': 0.3341992199420929, 'rewards/total_composite/mean': 0.17207136750221252, 'rewards/total_composite/std': 0.16055873036384583, 'reward': 0.17207136750221252, 'reward_std': 0.16055871546268463, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1683012694120407, 'sampling/sampling_logp_difference/max': 2.6271228790283203, 'sampling/importance_sampling_ratio/min': 0.07228614389896393, 'sampling/importance_sampling_ratio/mean': 1.000517725944519, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6649110615253448, 'clip_ratio/low_mean': 0.05331201106309891, 'clip_ratio/low_min': 0.05331201106309891, 'clip_ratio/high_mean': 0.07885371707379818, 'clip_ratio/high_max': 0.07885371707379818, 'clip_ratio/region_mean': 0.1321657281368971, 'reward_total_mean': 0.17207136750221252, 'reward_meter_mean': 0.4819783866405487, 'reward_meter_std': 0.3900616765022278, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.09258200973272324, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7841250896453857, 'reward_repeat_penalty_std': 0.3230384290218353, 'reward_near_duplicate_penalty_mean': 0.9610974192619324, 'reward_near_duplicate_penalty_std': 0.01954832673072815, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.816798746585846, 'reward_distinct_2_std': 0.3341992199420929, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.17207136750221252, 'reward_total_composite_std': 0.16055873036384583, 'epoch': 0.03}

 20%|██        | 672/3300 [2:39:54<8:36:12, 11.79s/it]INFO 04-12 13:26:06 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 673/3300 [2:40:03<8:01:13, 10.99s/it]2026-04-12 13:26:15,755 | INFO | train_grpo_train | metrics_logged mode=train step=673

                                                      
{'loss': 0.0454, 'grad_norm': 14.057157516479492, 'learning_rate': 7.963636363636365e-06, 'num_tokens': 1352888.0, 'completions/mean_length': 46.125, 'completions/min_length': 42.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.125, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.8637489080429077, 'rewards/meter/std': 0.1764443814754486, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4587499797344208, 'rewards/judge_quality/std': 0.192831352353096, 'rewards/repeat_penalty/mean': 0.9530385136604309, 'rewards/repeat_penalty/std': 0.04393855109810829, 'rewards/near_duplicate_penalty/mean': 0.961021900177002, 'rewards/near_duplicate_penalty/std': 0.023450659587979317, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9912587404251099, 'rewards/distinct_2/std': 0.024724015966057777, 'rewards/total_composite/mean': 0.39170509576797485, 'rewards/total_composite/std': 0.17069067060947418, 'reward': 0.39170509576797485, 'reward_std': 0.17069067060947418, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13802801072597504, 'sampling/sampling_logp_difference/max': 3.609140634536743, 'sampling/importance_sampling_ratio/min': 0.027075106278061867, 'sampling/importance_sampling_ratio/mean': 1.0172617435455322, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6700945235788822, 'clip_ratio/low_mean': 0.060082800686359406, 'clip_ratio/low_min': 0.060082800686359406, 'clip_ratio/high_mean': 0.05822447966784239, 'clip_ratio/high_max': 0.05822447966784239, 'clip_ratio/region_mean': 0.1183072803542018, 'reward_total_mean': 0.39170509576797485, 'reward_meter_mean': 0.8637489080429077, 'reward_meter_std': 0.1764443814754486, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9530385136604309, 'reward_repeat_penalty_std': 0.04393855109810829, 'reward_near_duplicate_penalty_mean': 0.961021900177002, 'reward_near_duplicate_penalty_std': 0.023450659587979317, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9912587404251099, 'reward_distinct_2_std': 0.024724015966057777, 'reward_judge_quality_mean': 0.4587499797344208, 'reward_judge_quality_std': 0.192831352353096, 'reward_total_composite_mean': 0.39170509576797485, 'reward_total_composite_std': 0.17069067060947418, 'epoch': 0.03}

 20%|██        | 673/3300 [2:40:03<8:01:13, 10.99s/it]INFO 04-12 13:26:15 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 674/3300 [2:40:12<7:36:46, 10.44s/it]2026-04-12 13:26:24,898 | INFO | train_grpo_train | metrics_logged mode=train step=674

                                                      
{'loss': -0.0281, 'grad_norm': 17.77109146118164, 'learning_rate': 7.96060606060606e-06, 'num_tokens': 1354487.0, 'completions/mean_length': 34.875, 'completions/min_length': 29.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.875, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.6136876344680786, 'rewards/meter/std': 0.3237185776233673, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5550000071525574, 'rewards/judge_quality/std': 0.22790977358818054, 'rewards/repeat_penalty/mean': 0.9697059392929077, 'rewards/repeat_penalty/std': 0.04509860649704933, 'rewards/near_duplicate_penalty/mean': 0.978339672088623, 'rewards/near_duplicate_penalty/std': 0.027134040370583534, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9908424615859985, 'rewards/distinct_2/std': 0.025901345536112785, 'rewards/total_composite/mean': 0.38173848390579224, 'rewards/total_composite/std': 0.32607918977737427, 'reward': 0.38173848390579224, 'reward_std': 0.32607918977737427, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15165205299854279, 'sampling/sampling_logp_difference/max': 1.9108219146728516, 'sampling/importance_sampling_ratio/min': 0.14795872569084167, 'sampling/importance_sampling_ratio/mean': 1.01705801486969, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0126424580812454, 'clip_ratio/low_mean': 0.08050890080630779, 'clip_ratio/low_min': 0.08050890080630779, 'clip_ratio/high_mean': 0.03880772087723017, 'clip_ratio/high_max': 0.03880772087723017, 'clip_ratio/region_mean': 0.11931662168353796, 'reward_total_mean': 0.38173848390579224, 'reward_meter_mean': 0.6136876344680786, 'reward_meter_std': 0.3237185776233673, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9697059392929077, 'reward_repeat_penalty_std': 0.04509860649704933, 'reward_near_duplicate_penalty_mean': 0.978339672088623, 'reward_near_duplicate_penalty_std': 0.027134040370583534, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9908424615859985, 'reward_distinct_2_std': 0.025901345536112785, 'reward_judge_quality_mean': 0.5550000071525574, 'reward_judge_quality_std': 0.22790977358818054, 'reward_total_composite_mean': 0.38173848390579224, 'reward_total_composite_std': 0.32607918977737427, 'epoch': 0.03}

 20%|██        | 674/3300 [2:40:12<7:36:46, 10.44s/it]INFO 04-12 13:26:25 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 675/3300 [2:40:24<7:58:47, 10.94s/it]2026-04-12 13:26:37,028 | INFO | train_grpo_train | metrics_logged mode=train step=675

                                                      
{'loss': 0.0224, 'grad_norm': 5.370761871337891, 'learning_rate': 7.957575757575758e-06, 'num_tokens': 1358401.0, 'completions/mean_length': 279.25, 'completions/min_length': 262.0, 'completions/max_length': 298.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 279.25, 'completions/min_terminated_length': 262.0, 'completions/max_terminated_length': 298.0, 'rewards/meter/mean': 0.8288944363594055, 'rewards/meter/std': 0.19590497016906738, 'rewards/count_adherence/mean': 0.8999999761581421, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.2540968358516693, 'rewards/repeat_penalty/std': 0.13142986595630646, 'rewards/near_duplicate_penalty/mean': 0.8205145597457886, 'rewards/near_duplicate_penalty/std': 0.061271749436855316, 'rewards/opening_diversity/mean': 0.9943181872367859, 'rewards/opening_diversity/std': 0.016070615500211716, 'rewards/distinct_2/mean': 0.3660544157028198, 'rewards/distinct_2/std': 0.15977324545383453, 'rewards/total_composite/mean': 0.17865273356437683, 'rewards/total_composite/std': 0.09777577966451645, 'reward': 0.17865273356437683, 'reward_std': 0.09777577221393585, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09400346875190735, 'sampling/sampling_logp_difference/max': 3.8435425758361816, 'sampling/importance_sampling_ratio/min': 0.021417593583464622, 'sampling/importance_sampling_ratio/mean': 0.9996914863586426, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4531523250043392, 'clip_ratio/low_mean': 0.05183503497391939, 'clip_ratio/low_min': 0.05183503497391939, 'clip_ratio/high_mean': 0.03676219470798969, 'clip_ratio/high_max': 0.03676219470798969, 'clip_ratio/region_mean': 0.08859722968190908, 'reward_total_mean': 0.17865273356437683, 'reward_meter_mean': 0.8288944363594055, 'reward_meter_std': 0.19590497016906738, 'reward_count_adherence_mean': 0.8999999761581421, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.2540968358516693, 'reward_repeat_penalty_std': 0.13142986595630646, 'reward_near_duplicate_penalty_mean': 0.8205145597457886, 'reward_near_duplicate_penalty_std': 0.061271749436855316, 'reward_opening_diversity_mean': 0.9943181872367859, 'reward_opening_diversity_std': 0.016070615500211716, 'reward_distinct_2_mean': 0.3660544157028198, 'reward_distinct_2_std': 0.15977324545383453, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.17865273356437683, 'reward_total_composite_std': 0.09777577966451645, 'epoch': 0.03}

 20%|██        | 675/3300 [2:40:24<7:58:47, 10.94s/it]INFO 04-12 13:26:37 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 676/3300 [2:40:33<7:22:53, 10.13s/it]2026-04-12 13:26:45,247 | INFO | train_grpo_train | metrics_logged mode=train step=676

                                                      
{'loss': 0.0441, 'grad_norm': 11.891114234924316, 'learning_rate': 7.954545454545455e-06, 'num_tokens': 1359869.0, 'completions/mean_length': 28.5, 'completions/min_length': 25.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.5, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.9678075909614563, 'rewards/meter/std': 0.056856755167245865, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5175000429153442, 'rewards/judge_quality/std': 0.26868730783462524, 'rewards/repeat_penalty/mean': 0.7450793981552124, 'rewards/repeat_penalty/std': 0.01093420758843422, 'rewards/near_duplicate_penalty/mean': 0.9934391975402832, 'rewards/near_duplicate_penalty/std': 0.014578936621546745, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4934595823287964, 'rewards/total_composite/std': 0.2405196726322174, 'reward': 0.4934595823287964, 'reward_std': 0.2405196577310562, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1260237842798233, 'sampling/sampling_logp_difference/max': 1.4793071746826172, 'sampling/importance_sampling_ratio/min': 0.22779546678066254, 'sampling/importance_sampling_ratio/mean': 1.0029480457305908, 'sampling/importance_sampling_ratio/max': 1.9300601482391357, 'entropy': 0.7648178711533546, 'clip_ratio/low_mean': 0.08059035800397396, 'clip_ratio/low_min': 0.08059035800397396, 'clip_ratio/high_mean': 0.029464286286383867, 'clip_ratio/high_max': 0.029464286286383867, 'clip_ratio/region_mean': 0.11005464429035783, 'reward_total_mean': 0.4934595823287964, 'reward_meter_mean': 0.9678075909614563, 'reward_meter_std': 0.056856755167245865, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7450793981552124, 'reward_repeat_penalty_std': 0.01093420758843422, 'reward_near_duplicate_penalty_mean': 0.9934391975402832, 'reward_near_duplicate_penalty_std': 0.014578936621546745, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5175000429153442, 'reward_judge_quality_std': 0.26868730783462524, 'reward_total_composite_mean': 0.4934595823287964, 'reward_total_composite_std': 0.2405196726322174, 'epoch': 0.03}

 20%|██        | 676/3300 [2:40:33<7:22:53, 10.13s/it]INFO 04-12 13:26:45 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 677/3300 [2:40:41<7:01:32,  9.64s/it]2026-04-12 13:26:53,760 | INFO | train_grpo_train | metrics_logged mode=train step=677

                                                      
{'loss': 0.0772, 'grad_norm': 12.06702709197998, 'learning_rate': 7.951515151515152e-06, 'num_tokens': 1361385.0, 'completions/mean_length': 43.5, 'completions/min_length': 39.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.5, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.6745848655700684, 'rewards/meter/std': 0.4040341377258301, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6012499928474426, 'rewards/judge_quality/std': 0.2975345253944397, 'rewards/repeat_penalty/mean': 0.9609465003013611, 'rewards/repeat_penalty/std': 0.0487876832485199, 'rewards/near_duplicate_penalty/mean': 0.9680687785148621, 'rewards/near_duplicate_penalty/std': 0.03588702902197838, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9923076629638672, 'rewards/distinct_2/std': 0.021757129579782486, 'rewards/total_composite/mean': 0.4496162533760071, 'rewards/total_composite/std': 0.34201952815055847, 'reward': 0.4496162533760071, 'reward_std': 0.34201952815055847, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12411359697580338, 'sampling/sampling_logp_difference/max': 1.5364108085632324, 'sampling/importance_sampling_ratio/min': 0.21515193581581116, 'sampling/importance_sampling_ratio/mean': 0.9982262849807739, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7586450204253197, 'clip_ratio/low_mean': 0.07098861550912261, 'clip_ratio/low_min': 0.07098861550912261, 'clip_ratio/high_mean': 0.046543041709810495, 'clip_ratio/high_max': 0.046543041709810495, 'clip_ratio/region_mean': 0.1175316572189331, 'reward_total_mean': 0.4496162533760071, 'reward_meter_mean': 0.6745848655700684, 'reward_meter_std': 0.4040341377258301, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9609465003013611, 'reward_repeat_penalty_std': 0.0487876832485199, 'reward_near_duplicate_penalty_mean': 0.9680687785148621, 'reward_near_duplicate_penalty_std': 0.03588702902197838, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9923076629638672, 'reward_distinct_2_std': 0.021757129579782486, 'reward_judge_quality_mean': 0.6012499928474426, 'reward_judge_quality_std': 0.2975345253944397, 'reward_total_composite_mean': 0.4496162533760071, 'reward_total_composite_std': 0.34201952815055847, 'epoch': 0.03}

 21%|██        | 677/3300 [2:40:41<7:01:32,  9.64s/it]INFO 04-12 13:26:54 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 678/3300 [2:40:51<7:09:54,  9.84s/it]2026-04-12 13:27:04,055 | INFO | train_grpo_train | metrics_logged mode=train step=678

                                                      
{'loss': 0.047, 'grad_norm': 7.9541168212890625, 'learning_rate': 7.948484848484848e-06, 'num_tokens': 1364109.0, 'completions/mean_length': 161.5, 'completions/min_length': 130.0, 'completions/max_length': 183.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 161.5, 'completions/min_terminated_length': 130.0, 'completions/max_terminated_length': 183.0, 'rewards/meter/mean': 0.8053493499755859, 'rewards/meter/std': 0.322150856256485, 'rewards/count_adherence/mean': 0.9285714626312256, 'rewards/count_adherence/std': 0.07636035233736038, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.11649646610021591, 'rewards/repeat_penalty/mean': 0.6077289581298828, 'rewards/repeat_penalty/std': 0.251259446144104, 'rewards/near_duplicate_penalty/mean': 0.8977565765380859, 'rewards/near_duplicate_penalty/std': 0.05994274839758873, 'rewards/opening_diversity/mean': 0.9955357313156128, 'rewards/opening_diversity/std': 0.012626901268959045, 'rewards/distinct_2/mean': 0.6936469078063965, 'rewards/distinct_2/std': 0.20930396020412445, 'rewards/total_composite/mean': 0.18478071689605713, 'rewards/total_composite/std': 0.17019730806350708, 'reward': 0.18478071689605713, 'reward_std': 0.1701972782611847, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11972744762897491, 'sampling/sampling_logp_difference/max': 2.36962890625, 'sampling/importance_sampling_ratio/min': 0.09351541846990585, 'sampling/importance_sampling_ratio/mean': 0.9945900440216064, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.592000599950552, 'clip_ratio/low_mean': 0.0536336125805974, 'clip_ratio/low_min': 0.0536336125805974, 'clip_ratio/high_mean': 0.05580764450132847, 'clip_ratio/high_max': 0.05580764450132847, 'clip_ratio/region_mean': 0.10944125708192587, 'reward_total_mean': 0.18478071689605713, 'reward_meter_mean': 0.8053493499755859, 'reward_meter_std': 0.322150856256485, 'reward_count_adherence_mean': 0.9285714626312256, 'reward_count_adherence_std': 0.07636035233736038, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6077289581298828, 'reward_repeat_penalty_std': 0.251259446144104, 'reward_near_duplicate_penalty_mean': 0.8977565765380859, 'reward_near_duplicate_penalty_std': 0.05994274839758873, 'reward_opening_diversity_mean': 0.9955357313156128, 'reward_opening_diversity_std': 0.012626901268959045, 'reward_distinct_2_mean': 0.6936469078063965, 'reward_distinct_2_std': 0.20930396020412445, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.11649646610021591, 'reward_total_composite_mean': 0.18478071689605713, 'reward_total_composite_std': 0.17019730806350708, 'epoch': 0.03}

 21%|██        | 678/3300 [2:40:52<7:09:54,  9.84s/it]INFO 04-12 13:27:04 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 679/3300 [2:41:01<7:04:55,  9.73s/it]2026-04-12 13:27:13,521 | INFO | train_grpo_train | metrics_logged mode=train step=679

                                                      
{'loss': 0.0183, 'grad_norm': 11.858473777770996, 'learning_rate': 7.945454545454547e-06, 'num_tokens': 1365876.0, 'completions/mean_length': 50.875, 'completions/min_length': 48.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.875, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.6219387054443359, 'rewards/meter/std': 0.4098900258541107, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6387500166893005, 'rewards/judge_quality/std': 0.24718055129051208, 'rewards/repeat_penalty/mean': 0.9427800178527832, 'rewards/repeat_penalty/std': 0.0702410340309143, 'rewards/near_duplicate_penalty/mean': 0.9608965516090393, 'rewards/near_duplicate_penalty/std': 0.042752623558044434, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9801586866378784, 'rewards/distinct_2/std': 0.036883533000946045, 'rewards/total_composite/mean': 0.39213526248931885, 'rewards/total_composite/std': 0.33854520320892334, 'reward': 0.39213526248931885, 'reward_std': 0.33854517340660095, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12446413189172745, 'sampling/sampling_logp_difference/max': 1.531905174255371, 'sampling/importance_sampling_ratio/min': 0.2161235213279724, 'sampling/importance_sampling_ratio/mean': 0.9956213235855103, 'sampling/importance_sampling_ratio/max': 1.9025863409042358, 'entropy': 0.664982732385397, 'clip_ratio/low_mean': 0.07277948781847954, 'clip_ratio/low_min': 0.07277948781847954, 'clip_ratio/high_mean': 0.0642069336026907, 'clip_ratio/high_max': 0.0642069336026907, 'clip_ratio/region_mean': 0.13698642142117023, 'reward_total_mean': 0.39213526248931885, 'reward_meter_mean': 0.6219387054443359, 'reward_meter_std': 0.4098900258541107, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9427800178527832, 'reward_repeat_penalty_std': 0.0702410340309143, 'reward_near_duplicate_penalty_mean': 0.9608965516090393, 'reward_near_duplicate_penalty_std': 0.042752623558044434, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9801586866378784, 'reward_distinct_2_std': 0.036883533000946045, 'reward_judge_quality_mean': 0.6387500166893005, 'reward_judge_quality_std': 0.24718055129051208, 'reward_total_composite_mean': 0.39213526248931885, 'reward_total_composite_std': 0.33854520320892334, 'epoch': 0.03}

 21%|██        | 679/3300 [2:41:01<7:04:55,  9.73s/it]INFO 04-12 13:27:13 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 680/3300 [2:41:11<7:08:41,  9.82s/it]2026-04-12 13:27:23,549 | INFO | train_grpo_train | metrics_logged mode=train step=680

                                                      
{'loss': -0.0238, 'grad_norm': 11.923002243041992, 'learning_rate': 7.942424242424242e-06, 'num_tokens': 1367841.0, 'completions/mean_length': 89.625, 'completions/min_length': 75.0, 'completions/max_length': 102.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 89.625, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.6149957180023193, 'rewards/meter/std': 0.3780932128429413, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.1414213478565216, 'rewards/repeat_penalty/mean': 0.8234045505523682, 'rewards/repeat_penalty/std': 0.057999514043331146, 'rewards/near_duplicate_penalty/mean': 0.928970992565155, 'rewards/near_duplicate_penalty/std': 0.030884819105267525, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8857342600822449, 'rewards/distinct_2/std': 0.041356515139341354, 'rewards/total_composite/mean': 0.2690228521823883, 'rewards/total_composite/std': 0.20418933033943176, 'reward': 0.2690228521823883, 'reward_std': 0.20418934524059296, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15171462297439575, 'sampling/sampling_logp_difference/max': 3.026287078857422, 'sampling/importance_sampling_ratio/min': 0.04849536716938019, 'sampling/importance_sampling_ratio/mean': 1.004455804824829, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8101786896586418, 'clip_ratio/low_mean': 0.06674056965857744, 'clip_ratio/low_min': 0.06674056965857744, 'clip_ratio/high_mean': 0.08735445328056812, 'clip_ratio/high_max': 0.08735445328056812, 'clip_ratio/region_mean': 0.15409502293914557, 'reward_total_mean': 0.2690228521823883, 'reward_meter_mean': 0.6149957180023193, 'reward_meter_std': 0.3780932128429413, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8234045505523682, 'reward_repeat_penalty_std': 0.057999514043331146, 'reward_near_duplicate_penalty_mean': 0.928970992565155, 'reward_near_duplicate_penalty_std': 0.030884819105267525, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8857342600822449, 'reward_distinct_2_std': 0.041356515139341354, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.1414213478565216, 'reward_total_composite_mean': 0.2690228521823883, 'reward_total_composite_std': 0.20418933033943176, 'epoch': 0.03}

 21%|██        | 680/3300 [2:41:11<7:08:41,  9.82s/it]INFO 04-12 13:27:23 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 681/3300 [2:41:24<7:53:52, 10.86s/it]2026-04-12 13:27:36,829 | INFO | train_grpo_train | metrics_logged mode=train step=681

                                                      
{'loss': -0.0927, 'grad_norm': 2.6674611568450928, 'learning_rate': 7.93939393939394e-06, 'num_tokens': 1369439.0, 'completions/mean_length': 162.75, 'completions/min_length': 38.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 46.333335876464844, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.7847678065299988, 'rewards/meter/std': 0.3550890386104584, '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/lexical_plausibility/mean': 0.9615534543991089, 'rewards/lexical_plausibility/std': 0.1087433397769928, 'rewards/judge_quality/mean': 0.512499988079071, 'rewards/judge_quality/std': 0.31949737668037415, 'rewards/repeat_penalty/mean': 0.8113471269607544, 'rewards/repeat_penalty/std': 0.34022751450538635, 'rewards/near_duplicate_penalty/mean': 0.9766889810562134, 'rewards/near_duplicate_penalty/std': 0.033630695194005966, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.8647711277008057, 'rewards/distinct_2/std': 0.3503357172012329, 'rewards/total_composite/mean': 0.44961053133010864, 'rewards/total_composite/std': 0.362165629863739, 'reward': 0.44961053133010864, 'reward_std': 0.3621656000614166, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12778885662555695, 'sampling/sampling_logp_difference/max': 1.2149319648742676, 'sampling/importance_sampling_ratio/min': 0.2967301905155182, 'sampling/importance_sampling_ratio/mean': 1.0071533918380737, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.565909743309021, 'clip_ratio/low_mean': 0.046725478023290634, 'clip_ratio/low_min': 0.046725478023290634, 'clip_ratio/high_mean': 0.0532852578908205, 'clip_ratio/high_max': 0.0532852578908205, 'clip_ratio/region_mean': 0.10001073591411114, 'reward_total_mean': 0.44961053133010864, 'reward_meter_mean': 0.7847678065299988, 'reward_meter_std': 0.3550890386104584, '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_lexical_plausibility_mean': 0.9615534543991089, 'reward_lexical_plausibility_std': 0.1087433397769928, 'reward_repeat_penalty_mean': 0.8113471269607544, 'reward_repeat_penalty_std': 0.34022751450538635, 'reward_near_duplicate_penalty_mean': 0.9766889810562134, 'reward_near_duplicate_penalty_std': 0.033630695194005966, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.8647711277008057, 'reward_distinct_2_std': 0.3503357172012329, 'reward_judge_quality_mean': 0.512499988079071, 'reward_judge_quality_std': 0.31949737668037415, 'reward_total_composite_mean': 0.44961053133010864, 'reward_total_composite_std': 0.362165629863739, 'epoch': 0.03}

 21%|██        | 681/3300 [2:41:24<7:53:52, 10.86s/it]INFO 04-12 13:27:37 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 682/3300 [2:41:35<7:49:33, 10.76s/it]2026-04-12 13:27:47,371 | INFO | train_grpo_train | metrics_logged mode=train step=682

                                                      
{'loss': -0.022, 'grad_norm': 10.661027908325195, 'learning_rate': 7.936363636363637e-06, 'num_tokens': 1371316.0, 'completions/mean_length': 81.625, 'completions/min_length': 72.0, 'completions/max_length': 90.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 81.625, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 90.0, 'rewards/meter/mean': 0.6773897409439087, 'rewards/meter/std': 0.3524584472179413, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.2749999761581421, 'rewards/judge_quality/std': 0.11649646610021591, 'rewards/repeat_penalty/mean': 0.9228302836418152, 'rewards/repeat_penalty/std': 0.07274725288152695, 'rewards/near_duplicate_penalty/mean': 0.9561222195625305, 'rewards/near_duplicate_penalty/std': 0.036774892359972, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.963905930519104, 'rewards/distinct_2/std': 0.04218706116080284, 'rewards/total_composite/mean': 0.1977335810661316, 'rewards/total_composite/std': 0.12415219843387604, 'reward': 0.1977335810661316, 'reward_std': 0.12415219843387604, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13656394183635712, 'sampling/sampling_logp_difference/max': 1.4903523921966553, 'sampling/importance_sampling_ratio/min': 0.22529326379299164, 'sampling/importance_sampling_ratio/mean': 1.0086195468902588, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6780933775007725, 'clip_ratio/low_mean': 0.05089546926319599, 'clip_ratio/low_min': 0.05089546926319599, 'clip_ratio/high_mean': 0.09456607233732939, 'clip_ratio/high_max': 0.09456607233732939, 'clip_ratio/region_mean': 0.14546154160052538, 'reward_total_mean': 0.1977335810661316, 'reward_meter_mean': 0.6773897409439087, 'reward_meter_std': 0.3524584472179413, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9228302836418152, 'reward_repeat_penalty_std': 0.07274725288152695, 'reward_near_duplicate_penalty_mean': 0.9561222195625305, 'reward_near_duplicate_penalty_std': 0.036774892359972, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.963905930519104, 'reward_distinct_2_std': 0.04218706116080284, 'reward_judge_quality_mean': 0.2749999761581421, 'reward_judge_quality_std': 0.11649646610021591, 'reward_total_composite_mean': 0.1977335810661316, 'reward_total_composite_std': 0.12415219843387604, 'epoch': 0.03}

 21%|██        | 682/3300 [2:41:35<7:49:33, 10.76s/it]INFO 04-12 13:27:47 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 683/3300 [2:41:45<7:35:41, 10.45s/it]2026-04-12 13:27:57,087 | INFO | train_grpo_train | metrics_logged mode=train step=683

                                                      
{'loss': 0.0256, 'grad_norm': 17.77046775817871, 'learning_rate': 7.933333333333334e-06, 'num_tokens': 1372919.0, 'completions/mean_length': 48.375, 'completions/min_length': 44.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.375, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.840705394744873, 'rewards/meter/std': 0.18037928640842438, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.512499988079071, 'rewards/judge_quality/std': 0.18468119204044342, 'rewards/repeat_penalty/mean': 0.9711122512817383, 'rewards/repeat_penalty/std': 0.04365866258740425, 'rewards/near_duplicate_penalty/mean': 0.9910552501678467, 'rewards/near_duplicate_penalty/std': 0.012532477267086506, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9797008633613586, 'rewards/distinct_2/std': 0.037656061351299286, 'rewards/total_composite/mean': 0.43711957335472107, 'rewards/total_composite/std': 0.20955386757850647, 'reward': 0.43711957335472107, 'reward_std': 0.20955388247966766, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11673174798488617, 'sampling/sampling_logp_difference/max': 2.4938652515411377, 'sampling/importance_sampling_ratio/min': 0.08259011805057526, 'sampling/importance_sampling_ratio/mean': 1.0123649835586548, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5816117040812969, 'clip_ratio/low_mean': 0.06230353470891714, 'clip_ratio/low_min': 0.06230353470891714, 'clip_ratio/high_mean': 0.03842670004814863, 'clip_ratio/high_max': 0.03842670004814863, 'clip_ratio/region_mean': 0.10073023475706577, 'reward_total_mean': 0.43711957335472107, 'reward_meter_mean': 0.840705394744873, 'reward_meter_std': 0.18037928640842438, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9711122512817383, 'reward_repeat_penalty_std': 0.04365866258740425, 'reward_near_duplicate_penalty_mean': 0.9910552501678467, 'reward_near_duplicate_penalty_std': 0.012532477267086506, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9797008633613586, 'reward_distinct_2_std': 0.037656061351299286, 'reward_judge_quality_mean': 0.512499988079071, 'reward_judge_quality_std': 0.18468119204044342, 'reward_total_composite_mean': 0.43711957335472107, 'reward_total_composite_std': 0.20955386757850647, 'epoch': 0.03}

 21%|██        | 683/3300 [2:41:45<7:35:41, 10.45s/it]INFO 04-12 13:27:57 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 684/3300 [2:41:54<7:20:43, 10.11s/it]2026-04-12 13:28:06,402 | INFO | train_grpo_train | metrics_logged mode=train step=684

                                                      
{'loss': 0.0906, 'grad_norm': 12.437556266784668, 'learning_rate': 7.930303030303031e-06, 'num_tokens': 1374680.0, 'completions/mean_length': 53.125, 'completions/min_length': 42.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.125, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.35926711559295654, 'rewards/meter/std': 0.33269840478897095, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6387500166893005, 'rewards/judge_quality/std': 0.3224211037158966, 'rewards/repeat_penalty/mean': 0.9056887626647949, 'rewards/repeat_penalty/std': 0.07590774446725845, 'rewards/near_duplicate_penalty/mean': 0.935394287109375, 'rewards/near_duplicate_penalty/std': 0.0471920445561409, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9673083424568176, 'rewards/distinct_2/std': 0.047377511858940125, 'rewards/total_composite/mean': 0.22235041856765747, 'rewards/total_composite/std': 0.24022987484931946, 'reward': 0.22235041856765747, 'reward_std': 0.24022985994815826, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14882268011569977, 'sampling/sampling_logp_difference/max': 2.0587077140808105, 'sampling/importance_sampling_ratio/min': 0.24324600398540497, 'sampling/importance_sampling_ratio/mean': 1.0149168968200684, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8003500439226627, 'clip_ratio/low_mean': 0.1035881545394659, 'clip_ratio/low_min': 0.1035881545394659, 'clip_ratio/high_mean': 0.021844115341082215, 'clip_ratio/high_max': 0.021844115341082215, 'clip_ratio/region_mean': 0.12543226988054812, 'reward_total_mean': 0.22235041856765747, 'reward_meter_mean': 0.35926711559295654, 'reward_meter_std': 0.33269840478897095, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9056887626647949, 'reward_repeat_penalty_std': 0.07590774446725845, 'reward_near_duplicate_penalty_mean': 0.935394287109375, 'reward_near_duplicate_penalty_std': 0.0471920445561409, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9673083424568176, 'reward_distinct_2_std': 0.047377511858940125, 'reward_judge_quality_mean': 0.6387500166893005, 'reward_judge_quality_std': 0.3224211037158966, 'reward_total_composite_mean': 0.22235041856765747, 'reward_total_composite_std': 0.24022987484931946, 'epoch': 0.03}

 21%|██        | 684/3300 [2:41:54<7:20:43, 10.11s/it]INFO 04-12 13:28:06 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 685/3300 [2:42:04<7:18:29, 10.06s/it]2026-04-12 13:28:16,355 | INFO | train_grpo_train | metrics_logged mode=train step=685

                                                      
{'loss': 0.0211, 'grad_norm': 10.001082420349121, 'learning_rate': 7.927272727272729e-06, 'num_tokens': 1377110.0, 'completions/mean_length': 108.75, 'completions/min_length': 103.0, 'completions/max_length': 119.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 108.75, 'completions/min_terminated_length': 103.0, 'completions/max_terminated_length': 119.0, 'rewards/meter/mean': 0.6494002342224121, 'rewards/meter/std': 0.3196577727794647, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.7960606217384338, 'rewards/repeat_penalty/std': 0.1494201123714447, 'rewards/near_duplicate_penalty/mean': 0.9417764544487, 'rewards/near_duplicate_penalty/std': 0.02592436410486698, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.8607696294784546, 'rewards/distinct_2/std': 0.11884584277868271, 'rewards/total_composite/mean': 0.25066906213760376, 'rewards/total_composite/std': 0.12427356094121933, 'reward': 0.25066906213760376, 'reward_std': 0.12427355349063873, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13754647970199585, 'sampling/sampling_logp_difference/max': 2.2427010536193848, 'sampling/importance_sampling_ratio/min': 0.10617134720087051, 'sampling/importance_sampling_ratio/mean': 0.997316300868988, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6303588226437569, 'clip_ratio/low_mean': 0.07756581623107195, 'clip_ratio/low_min': 0.07756581623107195, 'clip_ratio/high_mean': 0.04796245601028204, 'clip_ratio/high_max': 0.04796245601028204, 'clip_ratio/region_mean': 0.125528272241354, 'reward_total_mean': 0.25066906213760376, 'reward_meter_mean': 0.6494002342224121, 'reward_meter_std': 0.3196577727794647, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7960606217384338, 'reward_repeat_penalty_std': 0.1494201123714447, 'reward_near_duplicate_penalty_mean': 0.9417764544487, 'reward_near_duplicate_penalty_std': 0.02592436410486698, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.8607696294784546, 'reward_distinct_2_std': 0.11884584277868271, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.25066906213760376, 'reward_total_composite_std': 0.12427356094121933, 'epoch': 0.03}

 21%|██        | 685/3300 [2:42:04<7:18:29, 10.06s/it]INFO 04-12 13:28:16 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 686/3300 [2:42:14<7:21:47, 10.14s/it]2026-04-12 13:28:26,681 | INFO | train_grpo_train | metrics_logged mode=train step=686

                                                      
{'loss': -0.0143, 'grad_norm': 7.473944664001465, 'learning_rate': 7.924242424242426e-06, 'num_tokens': 1379655.0, 'completions/mean_length': 143.125, 'completions/min_length': 113.0, 'completions/max_length': 161.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 143.125, 'completions/min_terminated_length': 113.0, 'completions/max_terminated_length': 161.0, 'rewards/meter/mean': 0.8714023232460022, 'rewards/meter/std': 0.18332542479038239, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.10690449178218842, 'rewards/repeat_penalty/mean': 0.28070366382598877, 'rewards/repeat_penalty/std': 0.24587886035442352, 'rewards/near_duplicate_penalty/mean': 0.7621680498123169, 'rewards/near_duplicate_penalty/std': 0.15426813066005707, 'rewards/opening_diversity/mean': 0.800000011920929, 'rewards/opening_diversity/std': 0.21876275539398193, 'rewards/distinct_2/mean': 0.3956928849220276, 'rewards/distinct_2/std': 0.2787923514842987, 'rewards/total_composite/mean': 0.19202126562595367, 'rewards/total_composite/std': 0.13257306814193726, 'reward': 0.19202126562595367, 'reward_std': 0.13257306814193726, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09486376494169235, 'sampling/sampling_logp_difference/max': 2.607607841491699, 'sampling/importance_sampling_ratio/min': 0.07371066510677338, 'sampling/importance_sampling_ratio/mean': 1.0006015300750732, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.48386170342564583, 'clip_ratio/low_mean': 0.04504252364858985, 'clip_ratio/low_min': 0.04504252364858985, 'clip_ratio/high_mean': 0.03957773372530937, 'clip_ratio/high_max': 0.03957773372530937, 'clip_ratio/region_mean': 0.08462025737389922, 'reward_total_mean': 0.19202126562595367, 'reward_meter_mean': 0.8714023232460022, 'reward_meter_std': 0.18332542479038239, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.28070366382598877, 'reward_repeat_penalty_std': 0.24587886035442352, 'reward_near_duplicate_penalty_mean': 0.7621680498123169, 'reward_near_duplicate_penalty_std': 0.15426813066005707, 'reward_opening_diversity_mean': 0.800000011920929, 'reward_opening_diversity_std': 0.21876275539398193, 'reward_distinct_2_mean': 0.3956928849220276, 'reward_distinct_2_std': 0.2787923514842987, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.10690449178218842, 'reward_total_composite_mean': 0.19202126562595367, 'reward_total_composite_std': 0.13257306814193726, 'epoch': 0.03}

 21%|██        | 686/3300 [2:42:14<7:21:47, 10.14s/it]INFO 04-12 13:28:26 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 687/3300 [2:42:24<7:15:45, 10.01s/it]2026-04-12 13:28:36,377 | INFO | train_grpo_train | metrics_logged mode=train step=687

                                                      
{'loss': 0.0115, 'grad_norm': 16.36661148071289, 'learning_rate': 7.921212121212122e-06, 'num_tokens': 1381198.0, 'completions/mean_length': 48.875, 'completions/min_length': 45.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.875, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.6924780011177063, 'rewards/meter/std': 0.34872353076934814, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.16690458357334137, 'rewards/repeat_penalty/mean': 0.9573004245758057, 'rewards/repeat_penalty/std': 0.04531080275774002, 'rewards/near_duplicate_penalty/mean': 0.9760246276855469, 'rewards/near_duplicate_penalty/std': 0.025973476469516754, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9807209968566895, 'rewards/distinct_2/std': 0.0357571542263031, 'rewards/total_composite/mean': 0.2855433225631714, 'rewards/total_composite/std': 0.20969447493553162, 'reward': 0.2855433225631714, 'reward_std': 0.20969446003437042, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16747529804706573, 'sampling/sampling_logp_difference/max': 2.515855312347412, 'sampling/importance_sampling_ratio/min': 0.08079377561807632, 'sampling/importance_sampling_ratio/mean': 0.9764614701271057, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8049458265304565, 'clip_ratio/low_mean': 0.07250000163912773, 'clip_ratio/low_min': 0.07250000163912773, 'clip_ratio/high_mean': 0.09820890054106712, 'clip_ratio/high_max': 0.09820890054106712, 'clip_ratio/region_mean': 0.17070890218019485, 'reward_total_mean': 0.2855433225631714, 'reward_meter_mean': 0.6924780011177063, 'reward_meter_std': 0.34872353076934814, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9573004245758057, 'reward_repeat_penalty_std': 0.04531080275774002, 'reward_near_duplicate_penalty_mean': 0.9760246276855469, 'reward_near_duplicate_penalty_std': 0.025973476469516754, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9807209968566895, 'reward_distinct_2_std': 0.0357571542263031, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.16690458357334137, 'reward_total_composite_mean': 0.2855433225631714, 'reward_total_composite_std': 0.20969447493553162, 'epoch': 0.03}

 21%|██        | 687/3300 [2:42:24<7:15:45, 10.01s/it]INFO 04-12 13:28:36 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 688/3300 [2:42:34<7:14:50,  9.99s/it]2026-04-12 13:28:46,323 | INFO | train_grpo_train | metrics_logged mode=train step=688

                                                      
{'loss': 0.0661, 'grad_norm': 14.963428497314453, 'learning_rate': 7.918181818181819e-06, 'num_tokens': 1382905.0, 'completions/mean_length': 48.375, 'completions/min_length': 43.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.375, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.4290289282798767, 'rewards/meter/std': 0.39136409759521484, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5337499380111694, 'rewards/judge_quality/std': 0.17104199528694153, 'rewards/repeat_penalty/mean': 0.9784921407699585, 'rewards/repeat_penalty/std': 0.02166825532913208, 'rewards/near_duplicate_penalty/mean': 0.9784921407699585, 'rewards/near_duplicate_penalty/std': 0.02166825532913208, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.21057572960853577, 'rewards/total_composite/std': 0.17712515592575073, 'reward': 0.21057572960853577, 'reward_std': 0.17712515592575073, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16132418811321259, 'sampling/sampling_logp_difference/max': 1.9058542251586914, 'sampling/importance_sampling_ratio/min': 0.14869557321071625, 'sampling/importance_sampling_ratio/mean': 1.0063754320144653, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0045779198408127, 'clip_ratio/low_mean': 0.04261253890581429, 'clip_ratio/low_min': 0.04261253890581429, 'clip_ratio/high_mean': 0.07253559492528439, 'clip_ratio/high_max': 0.07253559492528439, 'clip_ratio/region_mean': 0.11514813383109868, 'reward_total_mean': 0.21057572960853577, 'reward_meter_mean': 0.4290289282798767, 'reward_meter_std': 0.39136409759521484, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9784921407699585, 'reward_repeat_penalty_std': 0.02166825532913208, 'reward_near_duplicate_penalty_mean': 0.9784921407699585, 'reward_near_duplicate_penalty_std': 0.02166825532913208, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5337499380111694, 'reward_judge_quality_std': 0.17104199528694153, 'reward_total_composite_mean': 0.21057572960853577, 'reward_total_composite_std': 0.17712515592575073, 'epoch': 0.03}

 21%|██        | 688/3300 [2:42:34<7:14:50,  9.99s/it]INFO 04-12 13:28:46 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 689/3300 [2:42:44<7:15:30, 10.01s/it]2026-04-12 13:28:56,373 | INFO | train_grpo_train | metrics_logged mode=train step=689

                                                      
{'loss': 0.0272, 'grad_norm': 7.877944469451904, 'learning_rate': 7.915151515151516e-06, 'num_tokens': 1385842.0, 'completions/mean_length': 159.125, 'completions/min_length': 145.0, 'completions/max_length': 176.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 159.125, 'completions/min_terminated_length': 145.0, 'completions/max_terminated_length': 176.0, 'rewards/meter/mean': 0.7990224957466125, 'rewards/meter/std': 0.2898417115211487, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3375000059604645, 'rewards/judge_quality/std': 0.08345229178667068, 'rewards/repeat_penalty/mean': 0.8136865496635437, 'rewards/repeat_penalty/std': 0.14370132982730865, 'rewards/near_duplicate_penalty/mean': 0.9447270035743713, 'rewards/near_duplicate_penalty/std': 0.029538163915276527, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8677237629890442, 'rewards/distinct_2/std': 0.1118897870182991, 'rewards/total_composite/mean': 0.26758864521980286, 'rewards/total_composite/std': 0.12349135428667068, 'reward': 0.26758864521980286, 'reward_std': 0.12349135428667068, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14809967577457428, 'sampling/sampling_logp_difference/max': 1.9510507583618164, 'sampling/importance_sampling_ratio/min': 0.14212465286254883, 'sampling/importance_sampling_ratio/mean': 1.0038474798202515, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7957801297307014, 'clip_ratio/low_mean': 0.06031614821404219, 'clip_ratio/low_min': 0.06031614821404219, 'clip_ratio/high_mean': 0.07692786864936352, 'clip_ratio/high_max': 0.07692786864936352, 'clip_ratio/region_mean': 0.1372440168634057, 'reward_total_mean': 0.26758864521980286, 'reward_meter_mean': 0.7990224957466125, 'reward_meter_std': 0.2898417115211487, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8136865496635437, 'reward_repeat_penalty_std': 0.14370132982730865, 'reward_near_duplicate_penalty_mean': 0.9447270035743713, 'reward_near_duplicate_penalty_std': 0.029538163915276527, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8677237629890442, 'reward_distinct_2_std': 0.1118897870182991, 'reward_judge_quality_mean': 0.3375000059604645, 'reward_judge_quality_std': 0.08345229178667068, 'reward_total_composite_mean': 0.26758864521980286, 'reward_total_composite_std': 0.12349135428667068, 'epoch': 0.03}

 21%|██        | 689/3300 [2:42:44<7:15:30, 10.01s/it]INFO 04-12 13:28:56 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 690/3300 [2:42:58<8:07:01, 11.20s/it]2026-04-12 13:29:10,341 | INFO | train_grpo_train | metrics_logged mode=train step=690

                                                      
{'loss': 0.0007, 'grad_norm': 3.4282939434051514, 'learning_rate': 7.912121212121213e-06, 'num_tokens': 1391062.0, 'completions/mean_length': 422.5, 'completions/min_length': 385.0, 'completions/max_length': 470.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 422.5, 'completions/min_terminated_length': 385.0, 'completions/max_terminated_length': 470.0, 'rewards/meter/mean': 0.8916245698928833, 'rewards/meter/std': 0.1614583283662796, 'rewards/count_adherence/mean': 0.9285714626312256, 'rewards/count_adherence/std': 0.03818017616868019, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.04020589590072632, 'rewards/repeat_penalty/std': 0.060685306787490845, 'rewards/near_duplicate_penalty/mean': 0.7095133066177368, 'rewards/near_duplicate_penalty/std': 0.10145653039216995, 'rewards/opening_diversity/mean': 0.9389951229095459, 'rewards/opening_diversity/std': 0.07091031968593597, 'rewards/distinct_2/mean': 0.06345638632774353, 'rewards/distinct_2/std': 0.09171286970376968, 'rewards/total_composite/mean': 0.1816328763961792, 'rewards/total_composite/std': 0.08169139176607132, 'reward': 0.1816328763961792, 'reward_std': 0.08169138431549072, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05731300637125969, 'sampling/sampling_logp_difference/max': 5.075409889221191, 'sampling/importance_sampling_ratio/min': 0.0062485248781740665, 'sampling/importance_sampling_ratio/mean': 1.0016580820083618, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.28396112471818924, 'clip_ratio/low_mean': 0.02942675177473575, 'clip_ratio/low_min': 0.02942675177473575, 'clip_ratio/high_mean': 0.023079270496964455, 'clip_ratio/high_max': 0.023079270496964455, 'clip_ratio/region_mean': 0.0525060222717002, 'reward_total_mean': 0.1816328763961792, 'reward_meter_mean': 0.8916245698928833, 'reward_meter_std': 0.1614583283662796, 'reward_count_adherence_mean': 0.9285714626312256, 'reward_count_adherence_std': 0.03818017616868019, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.04020589590072632, 'reward_repeat_penalty_std': 0.060685306787490845, 'reward_near_duplicate_penalty_mean': 0.7095133066177368, 'reward_near_duplicate_penalty_std': 0.10145653039216995, 'reward_opening_diversity_mean': 0.9389951229095459, 'reward_opening_diversity_std': 0.07091031968593597, 'reward_distinct_2_mean': 0.06345638632774353, 'reward_distinct_2_std': 0.09171286970376968, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.1816328763961792, 'reward_total_composite_std': 0.08169139176607132, 'epoch': 0.03}

 21%|██        | 690/3300 [2:42:58<8:07:01, 11.20s/it]INFO 04-12 13:29:10 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 691/3300 [2:43:06<7:31:51, 10.39s/it]2026-04-12 13:29:18,855 | INFO | train_grpo_train | metrics_logged mode=train step=691

                                                      
{'loss': 0.1545, 'grad_norm': 19.48782730102539, 'learning_rate': 7.909090909090909e-06, 'num_tokens': 1392662.0, 'completions/mean_length': 34.0, 'completions/min_length': 24.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.0, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.4723680913448334, 'rewards/meter/std': 0.4853588938713074, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7262499928474426, 'rewards/judge_quality/std': 0.2859039604663849, 'rewards/repeat_penalty/mean': 0.8004690408706665, 'rewards/repeat_penalty/std': 0.09692622721195221, 'rewards/near_duplicate_penalty/mean': 0.998526930809021, 'rewards/near_duplicate_penalty/std': 0.004166448023170233, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9893162250518799, 'rewards/distinct_2/std': 0.03021823801100254, 'rewards/total_composite/mean': 0.25151702761650085, 'rewards/total_composite/std': 0.36730679869651794, 'reward': 0.25151702761650085, 'reward_std': 0.36730679869651794, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14715968072414398, 'sampling/sampling_logp_difference/max': 2.48984432220459, 'sampling/importance_sampling_ratio/min': 0.08292287588119507, 'sampling/importance_sampling_ratio/mean': 1.0106021165847778, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6632404625415802, 'clip_ratio/low_mean': 0.06620554672554135, 'clip_ratio/low_min': 0.06620554672554135, 'clip_ratio/high_mean': 0.03504807688295841, 'clip_ratio/high_max': 0.03504807688295841, 'clip_ratio/region_mean': 0.10125362360849977, 'reward_total_mean': 0.25151702761650085, 'reward_meter_mean': 0.4723680913448334, 'reward_meter_std': 0.4853588938713074, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8004690408706665, 'reward_repeat_penalty_std': 0.09692622721195221, 'reward_near_duplicate_penalty_mean': 0.998526930809021, 'reward_near_duplicate_penalty_std': 0.004166448023170233, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9893162250518799, 'reward_distinct_2_std': 0.03021823801100254, 'reward_judge_quality_mean': 0.7262499928474426, 'reward_judge_quality_std': 0.2859039604663849, 'reward_total_composite_mean': 0.25151702761650085, 'reward_total_composite_std': 0.36730679869651794, 'epoch': 0.03}

 21%|██        | 691/3300 [2:43:06<7:31:51, 10.39s/it]INFO 04-12 13:29:19 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 692/3300 [2:43:15<7:14:21,  9.99s/it]2026-04-12 13:29:27,918 | INFO | train_grpo_train | metrics_logged mode=train step=692

                                                      
{'loss': 0.0825, 'grad_norm': 12.445843696594238, 'learning_rate': 7.906060606060608e-06, 'num_tokens': 1394428.0, 'completions/mean_length': 56.75, 'completions/min_length': 48.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.75, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.8535615801811218, 'rewards/meter/std': 0.24634356796741486, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5425000190734863, 'rewards/judge_quality/std': 0.23705033957958221, 'rewards/repeat_penalty/mean': 0.9866340160369873, 'rewards/repeat_penalty/std': 0.019210243597626686, 'rewards/near_duplicate_penalty/mean': 0.9866340160369873, 'rewards/near_duplicate_penalty/std': 0.019210243597626686, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.41610342264175415, 'rewards/total_composite/std': 0.28258654475212097, 'reward': 0.41610342264175415, 'reward_std': 0.28258654475212097, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15511003136634827, 'sampling/sampling_logp_difference/max': 2.0940399169921875, 'sampling/importance_sampling_ratio/min': 0.12318845838308334, 'sampling/importance_sampling_ratio/mean': 0.9979813694953918, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8518078625202179, 'clip_ratio/low_mean': 0.06950231455266476, 'clip_ratio/low_min': 0.06950231455266476, 'clip_ratio/high_mean': 0.08688047714531422, 'clip_ratio/high_max': 0.08688047714531422, 'clip_ratio/region_mean': 0.15638279169797897, 'reward_total_mean': 0.41610342264175415, 'reward_meter_mean': 0.8535615801811218, 'reward_meter_std': 0.24634356796741486, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9866340160369873, 'reward_repeat_penalty_std': 0.019210243597626686, 'reward_near_duplicate_penalty_mean': 0.9866340160369873, 'reward_near_duplicate_penalty_std': 0.019210243597626686, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5425000190734863, 'reward_judge_quality_std': 0.23705033957958221, 'reward_total_composite_mean': 0.41610342264175415, 'reward_total_composite_std': 0.28258654475212097, 'epoch': 0.03}

 21%|██        | 692/3300 [2:43:15<7:14:21,  9.99s/it]INFO 04-12 13:29:28 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 693/3300 [2:43:24<6:55:17,  9.56s/it]2026-04-12 13:29:36,461 | INFO | train_grpo_train | metrics_logged mode=train step=693

                                                      
{'loss': 0.1626, 'grad_norm': 16.988096237182617, 'learning_rate': 7.903030303030303e-06, 'num_tokens': 1395968.0, 'completions/mean_length': 37.5, 'completions/min_length': 32.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.5, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.588752806186676, 'rewards/meter/std': 0.38058748841285706, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.42499998211860657, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.9150839447975159, 'rewards/repeat_penalty/std': 0.0962492823600769, 'rewards/near_duplicate_penalty/mean': 0.9545110464096069, 'rewards/near_duplicate_penalty/std': 0.025570329278707504, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9575936794281006, 'rewards/distinct_2/std': 0.08631627261638641, 'rewards/total_composite/mean': 0.2420070767402649, 'rewards/total_composite/std': 0.152871772646904, 'reward': 0.2420070767402649, 'reward_std': 0.1528717577457428, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14583992958068848, 'sampling/sampling_logp_difference/max': 1.6880861520767212, 'sampling/importance_sampling_ratio/min': 0.18487299978733063, 'sampling/importance_sampling_ratio/mean': 1.003678560256958, 'sampling/importance_sampling_ratio/max': 1.8421077728271484, 'entropy': 0.7335262931883335, 'clip_ratio/low_mean': 0.0400134869851172, 'clip_ratio/low_min': 0.0400134869851172, 'clip_ratio/high_mean': 0.10827442817389965, 'clip_ratio/high_max': 0.10827442817389965, 'clip_ratio/region_mean': 0.14828791515901685, 'reward_total_mean': 0.2420070767402649, 'reward_meter_mean': 0.588752806186676, 'reward_meter_std': 0.38058748841285706, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9150839447975159, 'reward_repeat_penalty_std': 0.0962492823600769, 'reward_near_duplicate_penalty_mean': 0.9545110464096069, 'reward_near_duplicate_penalty_std': 0.025570329278707504, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9575936794281006, 'reward_distinct_2_std': 0.08631627261638641, 'reward_judge_quality_mean': 0.42499998211860657, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.2420070767402649, 'reward_total_composite_std': 0.152871772646904, 'epoch': 0.03}

 21%|██        | 693/3300 [2:43:24<6:55:17,  9.56s/it]INFO 04-12 13:29:36 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 694/3300 [2:43:32<6:41:09,  9.24s/it]2026-04-12 13:29:44,949 | INFO | train_grpo_train | metrics_logged mode=train step=694

                                                      
{'loss': 0.0104, 'grad_norm': 13.769287109375, 'learning_rate': 7.9e-06, 'num_tokens': 1397651.0, 'completions/mean_length': 48.375, 'completions/min_length': 43.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.375, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.3852972388267517, 'rewards/meter/std': 0.29132166504859924, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6725000143051147, 'rewards/judge_quality/std': 0.2666056752204895, 'rewards/repeat_penalty/mean': 0.8439086675643921, 'rewards/repeat_penalty/std': 0.13462969660758972, 'rewards/near_duplicate_penalty/mean': 0.923346757888794, 'rewards/near_duplicate_penalty/std': 0.04904133081436157, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9397153258323669, 'rewards/distinct_2/std': 0.06966220587491989, 'rewards/total_composite/mean': 0.2518223822116852, 'rewards/total_composite/std': 0.18078495562076569, 'reward': 0.2518223822116852, 'reward_std': 0.1807849407196045, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10397494584321976, 'sampling/sampling_logp_difference/max': 1.1548477411270142, 'sampling/importance_sampling_ratio/min': 0.31510552763938904, 'sampling/importance_sampling_ratio/mean': 1.008196234703064, 'sampling/importance_sampling_ratio/max': 1.869637131690979, 'entropy': 0.6907190605998039, 'clip_ratio/low_mean': 0.0661906199529767, 'clip_ratio/low_min': 0.0661906199529767, 'clip_ratio/high_mean': 0.07668428681790829, 'clip_ratio/high_max': 0.07668428681790829, 'clip_ratio/region_mean': 0.142874906770885, 'reward_total_mean': 0.2518223822116852, 'reward_meter_mean': 0.3852972388267517, 'reward_meter_std': 0.29132166504859924, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8439086675643921, 'reward_repeat_penalty_std': 0.13462969660758972, 'reward_near_duplicate_penalty_mean': 0.923346757888794, 'reward_near_duplicate_penalty_std': 0.04904133081436157, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9397153258323669, 'reward_distinct_2_std': 0.06966220587491989, 'reward_judge_quality_mean': 0.6725000143051147, 'reward_judge_quality_std': 0.2666056752204895, 'reward_total_composite_mean': 0.2518223822116852, 'reward_total_composite_std': 0.18078495562076569, 'epoch': 0.03}

 21%|██        | 694/3300 [2:43:32<6:41:09,  9.24s/it]INFO 04-12 13:29:45 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 695/3300 [2:43:47<7:55:10, 10.94s/it]2026-04-12 13:29:59,881 | INFO | train_grpo_train | metrics_logged mode=train step=695

                                                      
{'loss': -0.1019, 'grad_norm': 5.338735103607178, 'learning_rate': 7.896969696969698e-06, 'num_tokens': 1399378.0, 'completions/mean_length': 119.875, 'completions/min_length': 51.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 63.857147216796875, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.8705810308456421, 'rewards/meter/std': 0.2341153621673584, '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/lexical_plausibility/mean': 0.9654704332351685, 'rewards/lexical_plausibility/std': 0.09766443073749542, 'rewards/judge_quality/mean': 0.48374998569488525, 'rewards/judge_quality/std': 0.25961166620254517, 'rewards/repeat_penalty/mean': 0.9646810293197632, 'rewards/repeat_penalty/std': 0.018230050802230835, 'rewards/near_duplicate_penalty/mean': 0.9646810293197632, 'rewards/near_duplicate_penalty/std': 0.018230050802230835, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4178609848022461, 'rewards/total_composite/std': 0.2828451991081238, 'reward': 0.4178609848022461, 'reward_std': 0.2828451991081238, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11048547923564911, 'sampling/sampling_logp_difference/max': 1.5811209678649902, 'sampling/importance_sampling_ratio/min': 0.20574432611465454, 'sampling/importance_sampling_ratio/mean': 1.004176378250122, 'sampling/importance_sampling_ratio/max': 1.7473589181900024, 'entropy': 0.6982764676213264, 'clip_ratio/low_mean': 0.0417844308540225, 'clip_ratio/low_min': 0.0417844308540225, 'clip_ratio/high_mean': 0.074519376270473, 'clip_ratio/high_max': 0.074519376270473, 'clip_ratio/region_mean': 0.1163038071244955, 'reward_total_mean': 0.4178609848022461, 'reward_meter_mean': 0.8705810308456421, 'reward_meter_std': 0.2341153621673584, '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_lexical_plausibility_mean': 0.9654704332351685, 'reward_lexical_plausibility_std': 0.09766443073749542, 'reward_repeat_penalty_mean': 0.9646810293197632, 'reward_repeat_penalty_std': 0.018230050802230835, 'reward_near_duplicate_penalty_mean': 0.9646810293197632, 'reward_near_duplicate_penalty_std': 0.018230050802230835, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.48374998569488525, 'reward_judge_quality_std': 0.25961166620254517, 'reward_total_composite_mean': 0.4178609848022461, 'reward_total_composite_std': 0.2828451991081238, 'epoch': 0.03}

 21%|██        | 695/3300 [2:43:47<7:55:10, 10.94s/it]INFO 04-12 13:30:00 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 696/3300 [2:43:58<7:54:51, 10.94s/it]2026-04-12 13:30:10,849 | INFO | train_grpo_train | metrics_logged mode=train step=696

                                                      
{'loss': 0.033, 'grad_norm': 9.337010383605957, 'learning_rate': 7.893939393939395e-06, 'num_tokens': 1401497.0, 'completions/mean_length': 93.875, 'completions/min_length': 88.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 93.875, 'completions/min_terminated_length': 88.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.5079565644264221, 'rewards/meter/std': 0.40059345960617065, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9977678656578064, 'rewards/lexical_plausibility/std': 0.006313450634479523, 'rewards/judge_quality/mean': 0.4999999701976776, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.9242500066757202, 'rewards/repeat_penalty/std': 0.08778683096170425, 'rewards/near_duplicate_penalty/mean': 0.9711850881576538, 'rewards/near_duplicate_penalty/std': 0.025060100480914116, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9503452181816101, 'rewards/distinct_2/std': 0.0717897042632103, 'rewards/total_composite/mean': 0.23421421647071838, 'rewards/total_composite/std': 0.17414434254169464, 'reward': 0.23421421647071838, 'reward_std': 0.17414434254169464, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11903917044401169, 'sampling/sampling_logp_difference/max': 1.3152618408203125, 'sampling/importance_sampling_ratio/min': 0.2684040367603302, 'sampling/importance_sampling_ratio/mean': 1.0150032043457031, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8241725638508797, 'clip_ratio/low_mean': 0.06018863990902901, 'clip_ratio/low_min': 0.06018863990902901, 'clip_ratio/high_mean': 0.041672642808407545, 'clip_ratio/high_max': 0.041672642808407545, 'clip_ratio/region_mean': 0.10186128271743655, 'reward_total_mean': 0.23421421647071838, 'reward_meter_mean': 0.5079565644264221, 'reward_meter_std': 0.40059345960617065, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9977678656578064, 'reward_lexical_plausibility_std': 0.006313450634479523, 'reward_repeat_penalty_mean': 0.9242500066757202, 'reward_repeat_penalty_std': 0.08778683096170425, 'reward_near_duplicate_penalty_mean': 0.9711850881576538, 'reward_near_duplicate_penalty_std': 0.025060100480914116, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9503452181816101, 'reward_distinct_2_std': 0.0717897042632103, 'reward_judge_quality_mean': 0.4999999701976776, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.23421421647071838, 'reward_total_composite_std': 0.17414434254169464, 'epoch': 0.03}

 21%|██        | 696/3300 [2:43:58<7:54:51, 10.94s/it]INFO 04-12 13:30:11 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 697/3300 [2:44:13<8:41:31, 12.02s/it]2026-04-12 13:30:25,353 | INFO | train_grpo_train | metrics_logged mode=train step=697

                                                      
{'loss': -0.0729, 'grad_norm': 4.8107686042785645, 'learning_rate': 7.89090909090909e-06, 'num_tokens': 1403186.0, 'completions/mean_length': 111.125, 'completions/min_length': 50.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 53.85714340209961, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.7558977007865906, 'rewards/meter/std': 0.2821967303752899, '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/lexical_plausibility/mean': 0.966004490852356, 'rewards/lexical_plausibility/std': 0.07419652491807938, 'rewards/judge_quality/mean': 0.5512499809265137, 'rewards/judge_quality/std': 0.3296508193016052, 'rewards/repeat_penalty/mean': 0.9656820297241211, 'rewards/repeat_penalty/std': 0.057992201298475266, 'rewards/near_duplicate_penalty/mean': 0.9748162627220154, 'rewards/near_duplicate_penalty/std': 0.03587019070982933, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9898927807807922, 'rewards/distinct_2/std': 0.027032768353819847, 'rewards/total_composite/mean': 0.4203924536705017, 'rewards/total_composite/std': 0.3291851878166199, 'reward': 0.4203924536705017, 'reward_std': 0.32918521761894226, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1522277295589447, 'sampling/sampling_logp_difference/max': 2.0605649948120117, 'sampling/importance_sampling_ratio/min': 0.12738198041915894, 'sampling/importance_sampling_ratio/mean': 1.0011069774627686, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7282736003398895, 'clip_ratio/low_mean': 0.09063034225255251, 'clip_ratio/low_min': 0.09063034225255251, 'clip_ratio/high_mean': 0.04279807023704052, 'clip_ratio/high_max': 0.04279807023704052, 'clip_ratio/region_mean': 0.13342841248959303, 'reward_total_mean': 0.4203924536705017, 'reward_meter_mean': 0.7558977007865906, 'reward_meter_std': 0.2821967303752899, '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_lexical_plausibility_mean': 0.966004490852356, 'reward_lexical_plausibility_std': 0.07419652491807938, 'reward_repeat_penalty_mean': 0.9656820297241211, 'reward_repeat_penalty_std': 0.057992201298475266, 'reward_near_duplicate_penalty_mean': 0.9748162627220154, 'reward_near_duplicate_penalty_std': 0.03587019070982933, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9898927807807922, 'reward_distinct_2_std': 0.027032768353819847, 'reward_judge_quality_mean': 0.5512499809265137, 'reward_judge_quality_std': 0.3296508193016052, 'reward_total_composite_mean': 0.4203924536705017, 'reward_total_composite_std': 0.3291851878166199, 'epoch': 0.03}

 21%|██        | 697/3300 [2:44:13<8:41:31, 12.02s/it]INFO 04-12 13:30:25 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 698/3300 [2:44:26<8:59:36, 12.44s/it]2026-04-12 13:30:38,782 | INFO | train_grpo_train | metrics_logged mode=train step=698

                                                      
{'loss': 0.0637, 'grad_norm': 4.015619277954102, 'learning_rate': 7.88787878787879e-06, 'num_tokens': 1407255.0, 'completions/mean_length': 315.625, 'completions/min_length': 290.0, 'completions/max_length': 365.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 315.625, 'completions/min_terminated_length': 290.0, 'completions/max_terminated_length': 365.0, 'rewards/meter/mean': 0.8323097229003906, 'rewards/meter/std': 0.3074272572994232, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.08908708393573761, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3375000059604645, 'rewards/judge_quality/std': 0.08345229178667068, 'rewards/repeat_penalty/mean': 0.1499895453453064, 'rewards/repeat_penalty/std': 0.09942632913589478, 'rewards/near_duplicate_penalty/mean': 0.7662094235420227, 'rewards/near_duplicate_penalty/std': 0.10368543863296509, 'rewards/opening_diversity/mean': 0.8797348737716675, 'rewards/opening_diversity/std': 0.22495663166046143, 'rewards/distinct_2/mean': 0.21349364519119263, 'rewards/distinct_2/std': 0.1306503266096115, 'rewards/total_composite/mean': 0.2966635525226593, 'rewards/total_composite/std': 0.12691465020179749, 'reward': 0.2966635525226593, 'reward_std': 0.1269146353006363, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07019565999507904, 'sampling/sampling_logp_difference/max': 1.841604232788086, 'sampling/importance_sampling_ratio/min': 0.15856285393238068, 'sampling/importance_sampling_ratio/mean': 1.0085245370864868, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.42864049412310123, 'clip_ratio/low_mean': 0.01555402809754014, 'clip_ratio/low_min': 0.01555402809754014, 'clip_ratio/high_mean': 0.05781267490237951, 'clip_ratio/high_max': 0.05781267490237951, 'clip_ratio/region_mean': 0.07336670299991965, 'reward_total_mean': 0.2966635525226593, 'reward_meter_mean': 0.8323097229003906, 'reward_meter_std': 0.3074272572994232, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.08908708393573761, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.1499895453453064, 'reward_repeat_penalty_std': 0.09942632913589478, 'reward_near_duplicate_penalty_mean': 0.7662094235420227, 'reward_near_duplicate_penalty_std': 0.10368543863296509, 'reward_opening_diversity_mean': 0.8797348737716675, 'reward_opening_diversity_std': 0.22495663166046143, 'reward_distinct_2_mean': 0.21349364519119263, 'reward_distinct_2_std': 0.1306503266096115, 'reward_judge_quality_mean': 0.3375000059604645, 'reward_judge_quality_std': 0.08345229178667068, 'reward_total_composite_mean': 0.2966635525226593, 'reward_total_composite_std': 0.12691465020179749, 'epoch': 0.03}

 21%|██        | 698/3300 [2:44:26<8:59:36, 12.44s/it]INFO 04-12 13:30:39 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 699/3300 [2:44:35<8:05:49, 11.21s/it]2026-04-12 13:30:47,103 | INFO | train_grpo_train | metrics_logged mode=train step=699

                                                      
{'loss': 0.041, 'grad_norm': 10.690315246582031, 'learning_rate': 7.884848484848485e-06, 'num_tokens': 1409031.0, 'completions/mean_length': 54.0, 'completions/min_length': 52.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.7896865606307983, 'rewards/meter/std': 0.25810378789901733, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7337499856948853, 'rewards/judge_quality/std': 0.24401040375232697, 'rewards/repeat_penalty/mean': 0.9703421592712402, 'rewards/repeat_penalty/std': 0.02812577411532402, 'rewards/near_duplicate_penalty/mean': 0.9703421592712402, 'rewards/near_duplicate_penalty/std': 0.02812577411532402, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5779894590377808, 'rewards/total_composite/std': 0.293804407119751, 'reward': 0.5779894590377808, 'reward_std': 0.293804407119751, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1145641952753067, 'sampling/sampling_logp_difference/max': 2.3285369873046875, 'sampling/importance_sampling_ratio/min': 0.09743819385766983, 'sampling/importance_sampling_ratio/mean': 1.0050643682479858, 'sampling/importance_sampling_ratio/max': 1.9547998905181885, 'entropy': 0.7624307572841644, 'clip_ratio/low_mean': 0.07457103114575148, 'clip_ratio/low_min': 0.07457103114575148, 'clip_ratio/high_mean': 0.04112876579165459, 'clip_ratio/high_max': 0.04112876579165459, 'clip_ratio/region_mean': 0.11569979693740606, 'reward_total_mean': 0.5779894590377808, 'reward_meter_mean': 0.7896865606307983, 'reward_meter_std': 0.25810378789901733, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9703421592712402, 'reward_repeat_penalty_std': 0.02812577411532402, 'reward_near_duplicate_penalty_mean': 0.9703421592712402, 'reward_near_duplicate_penalty_std': 0.02812577411532402, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7337499856948853, 'reward_judge_quality_std': 0.24401040375232697, 'reward_total_composite_mean': 0.5779894590377808, 'reward_total_composite_std': 0.293804407119751, 'epoch': 0.03}

 21%|██        | 699/3300 [2:44:35<8:05:49, 11.21s/it]INFO 04-12 13:30:47 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 700/3300 [2:44:49<8:43:44, 12.09s/it]2026-04-12 13:31:01,242 | INFO | train_grpo_train | metrics_logged mode=train step=700

                                                      
{'loss': -0.0224, 'grad_norm': 3.9249844551086426, 'learning_rate': 7.881818181818182e-06, 'num_tokens': 1413897.0, 'completions/mean_length': 409.25, 'completions/min_length': 372.0, 'completions/max_length': 481.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 409.25, 'completions/min_terminated_length': 372.0, 'completions/max_terminated_length': 481.0, 'rewards/meter/mean': 0.8036543130874634, 'rewards/meter/std': 0.1369716227054596, 'rewards/count_adherence/mean': 0.9270833730697632, 'rewards/count_adherence/std': 0.06954358518123627, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.2629052698612213, 'rewards/repeat_penalty/std': 0.1960436850786209, 'rewards/near_duplicate_penalty/mean': 0.7825131416320801, 'rewards/near_duplicate_penalty/std': 0.14020468294620514, 'rewards/opening_diversity/mean': 0.9412775039672852, 'rewards/opening_diversity/std': 0.09353276342153549, 'rewards/distinct_2/mean': 0.32654058933258057, 'rewards/distinct_2/std': 0.20398451387882233, 'rewards/total_composite/mean': 0.08333609998226166, 'rewards/total_composite/std': 0.052965838462114334, 'reward': 0.08333609998226166, 'reward_std': 0.05296584218740463, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09280326962471008, 'sampling/sampling_logp_difference/max': 4.093554973602295, 'sampling/importance_sampling_ratio/min': 0.016679832711815834, 'sampling/importance_sampling_ratio/mean': 1.0056970119476318, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6075820289552212, 'clip_ratio/low_mean': 0.018030008301138878, 'clip_ratio/low_min': 0.018030008301138878, 'clip_ratio/high_mean': 0.06617697002366185, 'clip_ratio/high_max': 0.06617697002366185, 'clip_ratio/region_mean': 0.08420697832480073, 'reward_total_mean': 0.08333609998226166, 'reward_meter_mean': 0.8036543130874634, 'reward_meter_std': 0.1369716227054596, 'reward_count_adherence_mean': 0.9270833730697632, 'reward_count_adherence_std': 0.06954358518123627, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.2629052698612213, 'reward_repeat_penalty_std': 0.1960436850786209, 'reward_near_duplicate_penalty_mean': 0.7825131416320801, 'reward_near_duplicate_penalty_std': 0.14020468294620514, 'reward_opening_diversity_mean': 0.9412775039672852, 'reward_opening_diversity_std': 0.09353276342153549, 'reward_distinct_2_mean': 0.32654058933258057, 'reward_distinct_2_std': 0.20398451387882233, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.08333609998226166, 'reward_total_composite_std': 0.052965838462114334, 'epoch': 0.03}

 21%|██        | 700/3300 [2:44:49<8:43:44, 12.09s/it]INFO 04-12 13:31:01 [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 [00:11<01:02,  5.68s/it]

 23%|██▎       | 3/13 [00:22<01:20,  8.04s/it]

 31%|███       | 4/13 [00:34<01:24,  9.39s/it]

 38%|███▊      | 5/13 [00:45<01:21, 10.14s/it]

 46%|████▌     | 6/13 [00:58<01:15, 10.84s/it]

 54%|█████▍    | 7/13 [01:09<01:05, 10.99s/it]

 62%|██████▏   | 8/13 [01:20<00:55, 11.15s/it]

 69%|██████▉   | 9/13 [01:32<00:45, 11.30s/it]

 77%|███████▋  | 10/13 [01:43<00:33, 11.22s/it]

 85%|████████▍ | 11/13 [01:54<00:22, 11.20s/it]

 92%|█████████▏| 12/13 [02:05<00:11, 11.15s/it]

100%|██████████| 13/13 [02:18<00:00, 11.48s/it]2026-04-12 13:33:31,365 | INFO | train_grpo_train | metrics_logged mode=eval step=700

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 150.0697, 'eval_samples_per_second': 0.693, 'eval_steps_per_second': 0.087, 'eval_num_tokens': 1413897.0, 'eval_completions/mean_length': 280.25, 'eval_completions/min_length': 71.53846153846153, 'eval_completions/max_length': 512.0, 'eval_completions/clipped_ratio': 0.3173076923076923, 'eval_completions/mean_terminated_length': 175.25238623985877, 'eval_completions/min_terminated_length': 71.53846153846153, 'eval_completions/max_terminated_length': 330.7692307692308, 'eval_rewards/meter/mean': 0.4329166870850783, 'eval_rewards/meter/std': 0.38807074611003584, 'eval_rewards/count_adherence/mean': 0.8408197302084702, 'eval_rewards/count_adherence/std': 0.2695508392957541, 'eval_rewards/arabic_clean/mean': 0.7403846153846154, 'eval_rewards/arabic_clean/std': 0.42438613222195554, 'eval_rewards/lexical_plausibility/mean': 0.9337593087783227, 'eval_rewards/lexical_plausibility/std': 0.110670716716693, 'eval_rewards/judge_quality/mean': 0.27413461070794326, 'eval_rewards/judge_quality/std': 0.17943278298928186, 'eval_rewards/repeat_penalty/mean': 0.7714204421410193, 'eval_rewards/repeat_penalty/std': 0.2853775161963243, 'eval_rewards/near_duplicate_penalty/mean': 0.9268902356808002, 'eval_rewards/near_duplicate_penalty/std': 0.09547289031056258, 'eval_rewards/opening_diversity/mean': 0.9619753910945013, 'eval_rewards/opening_diversity/std': 0.084259062145765, 'eval_rewards/distinct_2/mean': 0.8305996197920579, 'eval_rewards/distinct_2/std': 0.27070970088243484, 'eval_rewards/total_composite/mean': 0.12682079466489646, 'eval_rewards/total_composite/std': 0.15041868560589278, 'eval_reward': 0.12682079466489646, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.0714514791392363, 'eval_sampling/sampling_logp_difference/max': 1.012208681840163, 'eval_sampling/importance_sampling_ratio/min': 0.37342764322574323, 'eval_sampling/importance_sampling_ratio/mean': 1.0172827794001653, 'eval_sampling/importance_sampling_ratio/max': 1.5328967112761278, 'eval_entropy': 0.8392358238880451, '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': 0.12682079466489646, 'eval_reward_meter_mean': 0.4329166870850783, 'eval_reward_meter_std': 0.38807074611003584, 'eval_reward_count_adherence_mean': 0.8408197302084702, 'eval_reward_count_adherence_std': 0.2695508392957541, 'eval_reward_arabic_clean_mean': 0.7403846153846154, 'eval_reward_arabic_clean_std': 0.42438613222195554, 'eval_reward_lexical_plausibility_mean': 0.9337593087783227, 'eval_reward_lexical_plausibility_std': 0.110670716716693, 'eval_reward_repeat_penalty_mean': 0.7714204421410193, 'eval_reward_repeat_penalty_std': 0.2853775161963243, 'eval_reward_near_duplicate_penalty_mean': 0.9268902356808002, 'eval_reward_near_duplicate_penalty_std': 0.09547289031056258, 'eval_reward_opening_diversity_mean': 0.9619753910945013, 'eval_reward_opening_diversity_std': 0.084259062145765, 'eval_reward_distinct_2_mean': 0.8305996197920579, 'eval_reward_distinct_2_std': 0.27070970088243484, 'eval_reward_judge_quality_mean': 0.27413461070794326, 'eval_reward_judge_quality_std': 0.17943278298928186, 'eval_reward_total_composite_mean': 0.12682079466489646, 'eval_reward_total_composite_std': 0.15041868560589278, 'epoch': 0.03}

 21%|██        | 700/3300 [2:47:19<8:43:44, 12.09s/it]

100%|██████████| 13/13 [02:18<00:00, 11.48s/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 [2:47:35<42:06:04, 58.32s/it]2026-04-12 13:33:47,431 | INFO | train_grpo_train | metrics_logged mode=train step=701

                                                       
{'loss': -0.1126, 'grad_norm': 4.728911399841309, 'learning_rate': 7.87878787878788e-06, 'num_tokens': 1415682.0, 'completions/mean_length': 121.125, 'completions/min_length': 59.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 65.28572082519531, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.4598010182380676, 'rewards/meter/std': 0.3762716054916382, '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/lexical_plausibility/mean': 0.9777243137359619, 'rewards/lexical_plausibility/std': 0.0630052387714386, 'rewards/judge_quality/mean': 0.3500000238418579, 'rewards/judge_quality/std': 0.2507132589817047, 'rewards/repeat_penalty/mean': 0.9004499912261963, 'rewards/repeat_penalty/std': 0.13890168070793152, 'rewards/near_duplicate_penalty/mean': 0.9440410733222961, 'rewards/near_duplicate_penalty/std': 0.0767950639128685, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9485930800437927, 'rewards/distinct_2/std': 0.07988756150007248, 'rewards/total_composite/mean': 0.12067333608865738, 'rewards/total_composite/std': 0.07071463763713837, 'reward': 0.12067333608865738, 'reward_std': 0.07071464508771896, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15905775129795074, 'sampling/sampling_logp_difference/max': 1.605569839477539, 'sampling/importance_sampling_ratio/min': 0.2007751166820526, 'sampling/importance_sampling_ratio/mean': 1.0211693048477173, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0422744303941727, 'clip_ratio/low_mean': 0.04112903214991093, 'clip_ratio/low_min': 0.04112903214991093, 'clip_ratio/high_mean': 0.09548537991940975, 'clip_ratio/high_max': 0.09548537991940975, 'clip_ratio/region_mean': 0.13661441206932068, 'reward_total_mean': 0.12067333608865738, 'reward_meter_mean': 0.4598010182380676, 'reward_meter_std': 0.3762716054916382, '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_lexical_plausibility_mean': 0.9777243137359619, 'reward_lexical_plausibility_std': 0.0630052387714386, 'reward_repeat_penalty_mean': 0.9004499912261963, 'reward_repeat_penalty_std': 0.13890168070793152, 'reward_near_duplicate_penalty_mean': 0.9440410733222961, 'reward_near_duplicate_penalty_std': 0.0767950639128685, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9485930800437927, 'reward_distinct_2_std': 0.07988756150007248, 'reward_judge_quality_mean': 0.3500000238418579, 'reward_judge_quality_std': 0.2507132589817047, 'reward_total_composite_mean': 0.12067333608865738, 'reward_total_composite_std': 0.07071463763713837, 'epoch': 0.03}

 21%|██        | 701/3300 [2:47:35<42:06:04, 58.32s/it]INFO 04-12 13:33:47 [block_pool.py:378] Successfully reset prefix cache

 21%|██▏       | 702/3300 [2:47:48<32:17:13, 44.74s/it]2026-04-12 13:34:00,492 | INFO | train_grpo_train | metrics_logged mode=train step=702

                                                       
{'loss': -0.0226, 'grad_norm': 5.830806732177734, 'learning_rate': 7.875757575757577e-06, 'num_tokens': 1417062.0, 'completions/mean_length': 100.5, 'completions/min_length': 37.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 41.71428680419922, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.4579460024833679, 'rewards/meter/std': 0.4362812042236328, '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/lexical_plausibility/mean': 0.9751198291778564, 'rewards/lexical_plausibility/std': 0.07037175446748734, 'rewards/judge_quality/mean': 0.4737499952316284, 'rewards/judge_quality/std': 0.30932819843292236, 'rewards/repeat_penalty/mean': 0.9724117517471313, 'rewards/repeat_penalty/std': 0.035894326865673065, 'rewards/near_duplicate_penalty/mean': 0.9724117517471313, 'rewards/near_duplicate_penalty/std': 0.035894326865673065, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2549290955066681, 'rewards/total_composite/std': 0.307390958070755, 'reward': 0.2549290955066681, 'reward_std': 0.3073909878730774, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16230902075767517, 'sampling/sampling_logp_difference/max': 2.509711503982544, 'sampling/importance_sampling_ratio/min': 0.08129168301820755, 'sampling/importance_sampling_ratio/mean': 1.0006777048110962, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6417121291160583, 'clip_ratio/low_mean': 0.06327936332672834, 'clip_ratio/low_min': 0.06327936332672834, 'clip_ratio/high_mean': 0.04319056170061231, 'clip_ratio/high_max': 0.04319056170061231, 'clip_ratio/region_mean': 0.10646992502734065, 'reward_total_mean': 0.2549290955066681, 'reward_meter_mean': 0.4579460024833679, 'reward_meter_std': 0.4362812042236328, '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_lexical_plausibility_mean': 0.9751198291778564, 'reward_lexical_plausibility_std': 0.07037175446748734, 'reward_repeat_penalty_mean': 0.9724117517471313, 'reward_repeat_penalty_std': 0.035894326865673065, 'reward_near_duplicate_penalty_mean': 0.9724117517471313, 'reward_near_duplicate_penalty_std': 0.035894326865673065, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4737499952316284, 'reward_judge_quality_std': 0.30932819843292236, 'reward_total_composite_mean': 0.2549290955066681, 'reward_total_composite_std': 0.307390958070755, 'epoch': 0.03}

 21%|██▏       | 702/3300 [2:47:48<32:17:13, 44.74s/it]INFO 04-12 13:34:00 [block_pool.py:378] Successfully reset prefix cache

 21%|██▏       | 703/3300 [2:48:02<25:32:27, 35.41s/it]2026-04-12 13:34:14,112 | INFO | train_grpo_train | metrics_logged mode=train step=703

                                                       
{'loss': -0.2044, 'grad_norm': 3.1238300800323486, 'learning_rate': 7.872727272727273e-06, 'num_tokens': 1419812.0, 'completions/mean_length': 257.75, 'completions/min_length': 156.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 173.0, 'completions/min_terminated_length': 156.0, 'completions/max_terminated_length': 197.0, 'rewards/meter/mean': 0.7507860064506531, 'rewards/meter/std': 0.42690497636795044, 'rewards/count_adherence/mean': 0.7083333134651184, 'rewards/count_adherence/std': 0.31810450553894043, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9290858507156372, 'rewards/lexical_plausibility/std': 0.1349487006664276, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.1505940705537796, 'rewards/repeat_penalty/mean': 0.9610195159912109, 'rewards/repeat_penalty/std': 0.027280161157250404, 'rewards/near_duplicate_penalty/mean': 0.9774726033210754, 'rewards/near_duplicate_penalty/std': 0.012866468168795109, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9830514192581177, 'rewards/distinct_2/std': 0.018757881596684456, 'rewards/total_composite/mean': 0.2014596313238144, 'rewards/total_composite/std': 0.16038274765014648, 'reward': 0.2014596313238144, 'reward_std': 0.16038274765014648, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14188724756240845, 'sampling/sampling_logp_difference/max': 2.0898890495300293, 'sampling/importance_sampling_ratio/min': 0.12370086461305618, 'sampling/importance_sampling_ratio/mean': 1.0204529762268066, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7533105984330177, 'clip_ratio/low_mean': 0.019508671015501022, 'clip_ratio/low_min': 0.019508671015501022, 'clip_ratio/high_mean': 0.09468829818069935, 'clip_ratio/high_max': 0.09468829818069935, 'clip_ratio/region_mean': 0.11419696919620037, 'reward_total_mean': 0.2014596313238144, 'reward_meter_mean': 0.7507860064506531, 'reward_meter_std': 0.42690497636795044, 'reward_count_adherence_mean': 0.7083333134651184, 'reward_count_adherence_std': 0.31810450553894043, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9290858507156372, 'reward_lexical_plausibility_std': 0.1349487006664276, 'reward_repeat_penalty_mean': 0.9610195159912109, 'reward_repeat_penalty_std': 0.027280161157250404, 'reward_near_duplicate_penalty_mean': 0.9774726033210754, 'reward_near_duplicate_penalty_std': 0.012866468168795109, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9830514192581177, 'reward_distinct_2_std': 0.018757881596684456, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.1505940705537796, 'reward_total_composite_mean': 0.2014596313238144, 'reward_total_composite_std': 0.16038274765014648, 'epoch': 0.03}

 21%|██▏       | 703/3300 [2:48:02<25:32:27, 35.41s/it]INFO 04-12 13:34:14 [block_pool.py:378] Successfully reset prefix cache

 21%|██▏       | 704/3300 [2:48:16<20:55:39, 29.02s/it]2026-04-12 13:34:28,241 | INFO | train_grpo_train | metrics_logged mode=train step=704

                                                       
{'loss': -0.2351, 'grad_norm': 3.3508265018463135, 'learning_rate': 7.86969696969697e-06, 'num_tokens': 1423139.0, 'completions/mean_length': 361.875, 'completions/min_length': 285.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 311.8333435058594, 'completions/min_terminated_length': 285.0, 'completions/max_terminated_length': 353.0, 'rewards/meter/mean': 0.4262239336967468, 'rewards/meter/std': 0.302864134311676, 'rewards/count_adherence/mean': 0.8409091234207153, 'rewards/count_adherence/std': 0.32867664098739624, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9639943838119507, 'rewards/lexical_plausibility/std': 0.091572105884552, 'rewards/judge_quality/mean': 0.23749999701976776, 'rewards/judge_quality/std': 0.11259915679693222, 'rewards/repeat_penalty/mean': 0.5871773958206177, 'rewards/repeat_penalty/std': 0.27425143122673035, 'rewards/near_duplicate_penalty/mean': 0.8991332054138184, 'rewards/near_duplicate_penalty/std': 0.06064100190997124, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.6544107794761658, 'rewards/distinct_2/std': 0.24710626900196075, 'rewards/total_composite/mean': 0.08113665878772736, 'rewards/total_composite/std': 0.06976733356714249, 'reward': 0.08113665878772736, 'reward_std': 0.06976732611656189, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11967301368713379, 'sampling/sampling_logp_difference/max': 3.156653881072998, 'sampling/importance_sampling_ratio/min': 0.14452998340129852, 'sampling/importance_sampling_ratio/mean': 1.013148307800293, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6142924204468727, 'clip_ratio/low_mean': 0.012719297781586647, 'clip_ratio/low_min': 0.012719297781586647, 'clip_ratio/high_mean': 0.07558081299066544, 'clip_ratio/high_max': 0.07558081299066544, 'clip_ratio/region_mean': 0.08830011077225208, 'reward_total_mean': 0.08113665878772736, 'reward_meter_mean': 0.4262239336967468, 'reward_meter_std': 0.302864134311676, 'reward_count_adherence_mean': 0.8409091234207153, 'reward_count_adherence_std': 0.32867664098739624, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9639943838119507, 'reward_lexical_plausibility_std': 0.091572105884552, 'reward_repeat_penalty_mean': 0.5871773958206177, 'reward_repeat_penalty_std': 0.27425143122673035, 'reward_near_duplicate_penalty_mean': 0.8991332054138184, 'reward_near_duplicate_penalty_std': 0.06064100190997124, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.6544107794761658, 'reward_distinct_2_std': 0.24710626900196075, 'reward_judge_quality_mean': 0.23749999701976776, 'reward_judge_quality_std': 0.11259915679693222, 'reward_total_composite_mean': 0.08113665878772736, 'reward_total_composite_std': 0.06976733356714249, 'epoch': 0.03}

 21%|██▏       | 704/3300 [2:48:16<20:55:39, 29.02s/it]INFO 04-12 13:34:28 [block_pool.py:378] Successfully reset prefix cache

 21%|██▏       | 705/3300 [2:48:29<17:35:21, 24.40s/it]2026-04-12 13:34:41,863 | INFO | train_grpo_train | metrics_logged mode=train step=705

                                                       
{'loss': -0.0388, 'grad_norm': 3.323040008544922, 'learning_rate': 7.866666666666667e-06, 'num_tokens': 1424714.0, 'completions/mean_length': 166.875, 'completions/min_length': 47.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 51.833335876464844, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.20770291984081268, 'rewards/meter/std': 0.3252376317977905, '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/lexical_plausibility/mean': 0.9348514080047607, 'rewards/lexical_plausibility/std': 0.12084727734327316, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.155264750123024, 'rewards/repeat_penalty/mean': 0.9069770574569702, 'rewards/repeat_penalty/std': 0.0928507000207901, 'rewards/near_duplicate_penalty/mean': 0.9542151093482971, 'rewards/near_duplicate_penalty/std': 0.04456498473882675, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9820234179496765, 'rewards/distinct_2/std': 0.033394016325473785, 'rewards/total_composite/mean': 0.07226188480854034, 'rewards/total_composite/std': 0.11443310976028442, 'reward': 0.07226188480854034, 'reward_std': 0.11443311721086502, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16576188802719116, 'sampling/sampling_logp_difference/max': 2.1529626846313477, 'sampling/importance_sampling_ratio/min': 0.11613956093788147, 'sampling/importance_sampling_ratio/mean': 1.0327199697494507, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0006302744150162, 'clip_ratio/low_mean': 0.07075394038110971, 'clip_ratio/low_min': 0.07075394038110971, 'clip_ratio/high_mean': 0.03052083309739828, 'clip_ratio/high_max': 0.03052083309739828, 'clip_ratio/region_mean': 0.101274773478508, 'reward_total_mean': 0.07226188480854034, 'reward_meter_mean': 0.20770291984081268, 'reward_meter_std': 0.3252376317977905, '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_lexical_plausibility_mean': 0.9348514080047607, 'reward_lexical_plausibility_std': 0.12084727734327316, 'reward_repeat_penalty_mean': 0.9069770574569702, 'reward_repeat_penalty_std': 0.0928507000207901, 'reward_near_duplicate_penalty_mean': 0.9542151093482971, 'reward_near_duplicate_penalty_std': 0.04456498473882675, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9820234179496765, 'reward_distinct_2_std': 0.033394016325473785, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.155264750123024, 'reward_total_composite_mean': 0.07226188480854034, 'reward_total_composite_std': 0.11443310976028442, 'epoch': 0.03}

 21%|██▏       | 705/3300 [2:48:29<17:35:21, 24.40s/it]INFO 04-12 13:34:42 [block_pool.py:378] Successfully reset prefix cache

 21%|██▏       | 706/3300 [2:48:37<14:00:33, 19.44s/it]2026-04-12 13:34:49,735 | INFO | train_grpo_train | metrics_logged mode=train step=706

                                                       
{'loss': 0.058, 'grad_norm': 16.509536743164062, 'learning_rate': 7.863636363636364e-06, 'num_tokens': 1426049.0, 'completions/mean_length': 34.875, 'completions/min_length': 33.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.875, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.7722452878952026, 'rewards/meter/std': 0.3245435953140259, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.7287236452102661, 'rewards/lexical_plausibility/std': 0.2502076327800751, 'rewards/judge_quality/mean': 0.6225000023841858, 'rewards/judge_quality/std': 0.3285357356071472, 'rewards/repeat_penalty/mean': 0.738068163394928, 'rewards/repeat_penalty/std': 0.033748287707567215, 'rewards/near_duplicate_penalty/mean': 0.9840909242630005, 'rewards/near_duplicate_penalty/std': 0.04499770700931549, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.45578867197036743, 'rewards/total_composite/std': 0.32262635231018066, 'reward': 0.45578867197036743, 'reward_std': 0.32262635231018066, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15979905426502228, 'sampling/sampling_logp_difference/max': 1.5976028442382812, 'sampling/importance_sampling_ratio/min': 0.20238107442855835, 'sampling/importance_sampling_ratio/mean': 1.034034013748169, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3599934056401253, 'clip_ratio/low_mean': 0.09168688952922821, 'clip_ratio/low_min': 0.09168688952922821, 'clip_ratio/high_mean': 0.06286605726927519, 'clip_ratio/high_max': 0.06286605726927519, 'clip_ratio/region_mean': 0.1545529467985034, 'reward_total_mean': 0.45578867197036743, 'reward_meter_mean': 0.7722452878952026, 'reward_meter_std': 0.3245435953140259, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.7287236452102661, 'reward_lexical_plausibility_std': 0.2502076327800751, 'reward_repeat_penalty_mean': 0.738068163394928, 'reward_repeat_penalty_std': 0.033748287707567215, 'reward_near_duplicate_penalty_mean': 0.9840909242630005, 'reward_near_duplicate_penalty_std': 0.04499770700931549, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6225000023841858, 'reward_judge_quality_std': 0.3285357356071472, 'reward_total_composite_mean': 0.45578867197036743, 'reward_total_composite_std': 0.32262635231018066, 'epoch': 0.03}

 21%|██▏       | 706/3300 [2:48:37<14:00:33, 19.44s/it]INFO 04-12 13:34:49 [block_pool.py:378] Successfully reset prefix cache

 21%|██▏       | 707/3300 [2:48:51<12:49:22, 17.80s/it]2026-04-12 13:35:03,709 | INFO | train_grpo_train | metrics_logged mode=train step=707

                                                       
{'loss': -0.0906, 'grad_norm': 2.5562710762023926, 'learning_rate': 7.860606060606062e-06, 'num_tokens': 1427778.0, 'completions/mean_length': 232.125, 'completions/min_length': 60.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 64.20000457763672, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.7281879186630249, 'rewards/meter/std': 0.3522622883319855, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.8970426321029663, 'rewards/lexical_plausibility/std': 0.15456870198249817, 'rewards/judge_quality/mean': 0.3087500035762787, 'rewards/judge_quality/std': 0.2881684899330139, 'rewards/repeat_penalty/mean': 0.878716230392456, 'rewards/repeat_penalty/std': 0.17452524602413177, 'rewards/near_duplicate_penalty/mean': 0.9691689014434814, 'rewards/near_duplicate_penalty/std': 0.03202034905552864, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9571457505226135, 'rewards/distinct_2/std': 0.06666997075080872, 'rewards/total_composite/mean': 0.1733127385377884, 'rewards/total_composite/std': 0.18337255716323853, 'reward': 0.1733127385377884, 'reward_std': 0.18337254226207733, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13388222455978394, 'sampling/sampling_logp_difference/max': 1.4603995084762573, 'sampling/importance_sampling_ratio/min': 0.23214350640773773, 'sampling/importance_sampling_ratio/mean': 1.0199592113494873, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6001369431614876, 'clip_ratio/low_mean': 0.014492753893136978, 'clip_ratio/low_min': 0.014492753893136978, 'clip_ratio/high_mean': 0.07541691791266203, 'clip_ratio/high_max': 0.07541691791266203, 'clip_ratio/region_mean': 0.08990967180579901, 'reward_total_mean': 0.1733127385377884, 'reward_meter_mean': 0.7281879186630249, 'reward_meter_std': 0.3522622883319855, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.8970426321029663, 'reward_lexical_plausibility_std': 0.15456870198249817, 'reward_repeat_penalty_mean': 0.878716230392456, 'reward_repeat_penalty_std': 0.17452524602413177, 'reward_near_duplicate_penalty_mean': 0.9691689014434814, 'reward_near_duplicate_penalty_std': 0.03202034905552864, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9571457505226135, 'reward_distinct_2_std': 0.06666997075080872, 'reward_judge_quality_mean': 0.3087500035762787, 'reward_judge_quality_std': 0.2881684899330139, 'reward_total_composite_mean': 0.1733127385377884, 'reward_total_composite_std': 0.18337255716323853, 'epoch': 0.03}

 21%|██▏       | 707/3300 [2:48:51<12:49:22, 17.80s/it]INFO 04-12 13:35:03 [block_pool.py:378] Successfully reset prefix cache

 21%|██▏       | 708/3300 [2:49:05<11:57:52, 16.62s/it]2026-04-12 13:35:17,563 | INFO | train_grpo_train | metrics_logged mode=train step=708

                                                       
{'loss': -0.1665, 'grad_norm': 2.5779967308044434, 'learning_rate': 7.857575757575759e-06, 'num_tokens': 1431047.0, 'completions/mean_length': 278.625, 'completions/min_length': 221.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 245.2857208251953, 'completions/min_terminated_length': 221.0, 'completions/max_terminated_length': 272.0, 'rewards/meter/mean': 0.8360555171966553, 'rewards/meter/std': 0.1717107594013214, 'rewards/count_adherence/mean': 0.805555522441864, 'rewards/count_adherence/std': 0.2832788825035095, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9661475419998169, 'rewards/lexical_plausibility/std': 0.09574912488460541, 'rewards/judge_quality/mean': 0.16249999403953552, 'rewards/judge_quality/std': 0.08345229923725128, 'rewards/repeat_penalty/mean': 0.5714437365531921, 'rewards/repeat_penalty/std': 0.21675854921340942, 'rewards/near_duplicate_penalty/mean': 0.8558686971664429, 'rewards/near_duplicate_penalty/std': 0.08912242203950882, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.6769595742225647, 'rewards/distinct_2/std': 0.17013928294181824, 'rewards/total_composite/mean': 0.12402010709047318, 'rewards/total_composite/std': 0.0695514976978302, 'reward': 0.12402010709047318, 'reward_std': 0.0695514902472496, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09929447621107101, 'sampling/sampling_logp_difference/max': 1.8009637594223022, 'sampling/importance_sampling_ratio/min': 0.16513966023921967, 'sampling/importance_sampling_ratio/mean': 1.012818694114685, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.469089113175869, 'clip_ratio/low_mean': 0.03810977004468441, 'clip_ratio/low_min': 0.03810977004468441, 'clip_ratio/high_mean': 0.042792328633368015, 'clip_ratio/high_max': 0.042792328633368015, 'clip_ratio/region_mean': 0.08090209867805243, 'reward_total_mean': 0.12402010709047318, 'reward_meter_mean': 0.8360555171966553, 'reward_meter_std': 0.1717107594013214, 'reward_count_adherence_mean': 0.805555522441864, 'reward_count_adherence_std': 0.2832788825035095, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9661475419998169, 'reward_lexical_plausibility_std': 0.09574912488460541, 'reward_repeat_penalty_mean': 0.5714437365531921, 'reward_repeat_penalty_std': 0.21675854921340942, 'reward_near_duplicate_penalty_mean': 0.8558686971664429, 'reward_near_duplicate_penalty_std': 0.08912242203950882, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.6769595742225647, 'reward_distinct_2_std': 0.17013928294181824, 'reward_judge_quality_mean': 0.16249999403953552, 'reward_judge_quality_std': 0.08345229923725128, 'reward_total_composite_mean': 0.12402010709047318, 'reward_total_composite_std': 0.0695514976978302, 'epoch': 0.03}

 21%|██▏       | 708/3300 [2:49:05<11:57:52, 16.62s/it]INFO 04-12 13:35:17 [block_pool.py:378] Successfully reset prefix cache

 21%|██▏       | 709/3300 [2:49:20<11:36:24, 16.13s/it]2026-04-12 13:35:32,547 | INFO | train_grpo_train | metrics_logged mode=train step=709

                                                       
{'loss': -0.2504, 'grad_norm': 2.9031665325164795, 'learning_rate': 7.854545454545454e-06, 'num_tokens': 1435307.0, 'completions/mean_length': 393.5, 'completions/min_length': 216.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 376.5714416503906, 'completions/min_terminated_length': 216.0, 'completions/max_terminated_length': 434.0, 'rewards/meter/mean': 0.6926078796386719, 'rewards/meter/std': 0.26816973090171814, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.38184911012649536, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9558064341545105, 'rewards/lexical_plausibility/std': 0.09961117058992386, 'rewards/judge_quality/mean': 0.13125000894069672, 'rewards/judge_quality/std': 0.0530330128967762, 'rewards/repeat_penalty/mean': 0.35398775339126587, 'rewards/repeat_penalty/std': 0.34143513441085815, 'rewards/near_duplicate_penalty/mean': 0.8688129186630249, 'rewards/near_duplicate_penalty/std': 0.08688666671514511, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.4254310131072998, 'rewards/distinct_2/std': 0.37536802887916565, 'rewards/total_composite/mean': 0.08600445091724396, 'rewards/total_composite/std': 0.05431681498885155, 'reward': 0.08600445091724396, 'reward_std': 0.05431681498885155, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1035347506403923, 'sampling/sampling_logp_difference/max': 2.9126782417297363, 'sampling/importance_sampling_ratio/min': 0.05433002859354019, 'sampling/importance_sampling_ratio/mean': 1.0033495426177979, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7287429347634315, 'clip_ratio/low_mean': 0.024561572819948196, 'clip_ratio/low_min': 0.024561572819948196, 'clip_ratio/high_mean': 0.05949160410091281, 'clip_ratio/high_max': 0.05949160410091281, 'clip_ratio/region_mean': 0.084053176920861, 'reward_total_mean': 0.08600445091724396, 'reward_meter_mean': 0.6926078796386719, 'reward_meter_std': 0.26816973090171814, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.38184911012649536, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9558064341545105, 'reward_lexical_plausibility_std': 0.09961117058992386, 'reward_repeat_penalty_mean': 0.35398775339126587, 'reward_repeat_penalty_std': 0.34143513441085815, 'reward_near_duplicate_penalty_mean': 0.8688129186630249, 'reward_near_duplicate_penalty_std': 0.08688666671514511, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.4254310131072998, 'reward_distinct_2_std': 0.37536802887916565, 'reward_judge_quality_mean': 0.13125000894069672, 'reward_judge_quality_std': 0.0530330128967762, 'reward_total_composite_mean': 0.08600445091724396, 'reward_total_composite_std': 0.05431681498885155, 'epoch': 0.03}

 21%|██▏       | 709/3300 [2:49:20<11:36:24, 16.13s/it]INFO 04-12 13:35:32 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 710/3300 [2:49:35<11:21:07, 15.78s/it]2026-04-12 13:35:47,511 | INFO | train_grpo_train | metrics_logged mode=train step=710

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 7.851515151515152e-06, 'num_tokens': 1437203.0, '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': 0.7577996253967285, 'rewards/meter/std': 0.2664344310760498, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.059608783572912216, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9957386255264282, 'rewards/lexical_plausibility/std': 0.012052967213094234, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.1225799024105072, 'rewards/repeat_penalty/std': 0.12920282781124115, 'rewards/near_duplicate_penalty/mean': 0.7553712725639343, 'rewards/near_duplicate_penalty/std': 0.09750039130449295, 'rewards/opening_diversity/mean': 0.9888392686843872, 'rewards/opening_diversity/std': 0.031567275524139404, 'rewards/distinct_2/mean': 0.17089681327342987, 'rewards/distinct_2/std': 0.17077679932117462, 'rewards/total_composite/mean': 0.08875250071287155, 'rewards/total_composite/std': 0.04935825243592262, 'reward': 0.08875250071287155, 'reward_std': 0.04935825616121292, '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': 0.08875250071287155, 'reward_meter_mean': 0.7577996253967285, 'reward_meter_std': 0.2664344310760498, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.059608783572912216, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9957386255264282, 'reward_lexical_plausibility_std': 0.012052967213094234, 'reward_repeat_penalty_mean': 0.1225799024105072, 'reward_repeat_penalty_std': 0.12920282781124115, 'reward_near_duplicate_penalty_mean': 0.7553712725639343, 'reward_near_duplicate_penalty_std': 0.09750039130449295, 'reward_opening_diversity_mean': 0.9888392686843872, 'reward_opening_diversity_std': 0.031567275524139404, 'reward_distinct_2_mean': 0.17089681327342987, 'reward_distinct_2_std': 0.17077679932117462, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.08875250071287155, 'reward_total_composite_std': 0.04935825243592262, 'epoch': 0.03}

 22%|██▏       | 710/3300 [2:49:35<11:21:07, 15.78s/it]INFO 04-12 13:35:47 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 711/3300 [2:49:50<11:14:24, 15.63s/it]2026-04-12 13:36:02,795 | INFO | train_grpo_train | metrics_logged mode=train step=711

                                                       
{'loss': -0.0751, 'grad_norm': 0.2859575152397156, 'learning_rate': 7.848484848484849e-06, 'num_tokens': 1440177.0, 'completions/mean_length': 500.75, 'completions/min_length': 425.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.75, 'completions/mean_terminated_length': 467.0, 'completions/min_terminated_length': 425.0, 'completions/max_terminated_length': 509.0, 'rewards/meter/mean': 0.9490606784820557, 'rewards/meter/std': 0.056531812995672226, 'rewards/count_adherence/mean': 0.9464285969734192, 'rewards/count_adherence/std': 0.05050763115286827, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.0008375698234885931, 'rewards/repeat_penalty/std': 0.0023690052330493927, 'rewards/near_duplicate_penalty/mean': 0.6331762075424194, 'rewards/near_duplicate_penalty/std': 0.09013549238443375, 'rewards/opening_diversity/mean': 0.8469989895820618, 'rewards/opening_diversity/std': 0.08452124893665314, 'rewards/distinct_2/mean': 0.0013736264081671834, 'rewards/distinct_2/std': 0.003885202342644334, 'rewards/total_composite/mean': 0.13455164432525635, 'rewards/total_composite/std': 0.008079596795141697, 'reward': 0.13455164432525635, 'reward_std': 0.008079597726464272, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.02639433741569519, 'sampling/sampling_logp_difference/max': 0.842078685760498, 'sampling/importance_sampling_ratio/min': 0.43081408739089966, 'sampling/importance_sampling_ratio/mean': 1.0005429983139038, 'sampling/importance_sampling_ratio/max': 1.7887802124023438, 'entropy': 0.05425930209457874, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.007167456205934286, 'clip_ratio/high_max': 0.007167456205934286, 'clip_ratio/region_mean': 0.007167456205934286, 'reward_total_mean': 0.13455164432525635, 'reward_meter_mean': 0.9490606784820557, 'reward_meter_std': 0.056531812995672226, 'reward_count_adherence_mean': 0.9464285969734192, 'reward_count_adherence_std': 0.05050763115286827, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.0008375698234885931, 'reward_repeat_penalty_std': 0.0023690052330493927, 'reward_near_duplicate_penalty_mean': 0.6331762075424194, 'reward_near_duplicate_penalty_std': 0.09013549238443375, 'reward_opening_diversity_mean': 0.8469989895820618, 'reward_opening_diversity_std': 0.08452124893665314, 'reward_distinct_2_mean': 0.0013736264081671834, 'reward_distinct_2_std': 0.003885202342644334, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.13455164432525635, 'reward_total_composite_std': 0.008079596795141697, 'epoch': 0.03}

 22%|██▏       | 711/3300 [2:49:50<11:14:24, 15.63s/it]INFO 04-12 13:36:03 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 712/3300 [2:50:04<10:45:49, 14.97s/it]2026-04-12 13:36:16,238 | INFO | train_grpo_train | metrics_logged mode=train step=712

                                                       
{'loss': -0.0866, 'grad_norm': 4.251730442047119, 'learning_rate': 7.845454545454546e-06, 'num_tokens': 1441841.0, 'completions/mean_length': 173.0, 'completions/min_length': 53.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 60.0, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.4674437642097473, 'rewards/meter/std': 0.4737145006656647, '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/lexical_plausibility/mean': 0.9053367376327515, 'rewards/lexical_plausibility/std': 0.17668111622333527, 'rewards/judge_quality/mean': 0.29374998807907104, 'rewards/judge_quality/std': 0.17204132676124573, 'rewards/repeat_penalty/mean': 0.9451378583908081, 'rewards/repeat_penalty/std': 0.041200459003448486, 'rewards/near_duplicate_penalty/mean': 0.9671118855476379, 'rewards/near_duplicate_penalty/std': 0.021453896537423134, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.977196216583252, 'rewards/distinct_2/std': 0.034176018089056015, 'rewards/total_composite/mean': 0.18675193190574646, 'rewards/total_composite/std': 0.19529974460601807, 'reward': 0.18675193190574646, 'reward_std': 0.19529974460601807, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13164152204990387, 'sampling/sampling_logp_difference/max': 1.9119737148284912, 'sampling/importance_sampling_ratio/min': 0.147788405418396, 'sampling/importance_sampling_ratio/mean': 1.0013500452041626, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.47536277025938034, 'clip_ratio/low_mean': 0.020233366638422012, 'clip_ratio/low_min': 0.020233366638422012, 'clip_ratio/high_mean': 0.0677831219509244, 'clip_ratio/high_max': 0.0677831219509244, 'clip_ratio/region_mean': 0.08801648858934641, 'reward_total_mean': 0.18675193190574646, 'reward_meter_mean': 0.4674437642097473, 'reward_meter_std': 0.4737145006656647, '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_lexical_plausibility_mean': 0.9053367376327515, 'reward_lexical_plausibility_std': 0.17668111622333527, 'reward_repeat_penalty_mean': 0.9451378583908081, 'reward_repeat_penalty_std': 0.041200459003448486, 'reward_near_duplicate_penalty_mean': 0.9671118855476379, 'reward_near_duplicate_penalty_std': 0.021453896537423134, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.977196216583252, 'reward_distinct_2_std': 0.034176018089056015, 'reward_judge_quality_mean': 0.29374998807907104, 'reward_judge_quality_std': 0.17204132676124573, 'reward_total_composite_mean': 0.18675193190574646, 'reward_total_composite_std': 0.19529974460601807, 'epoch': 0.03}

 22%|██▏       | 712/3300 [2:50:04<10:45:49, 14.97s/it]INFO 04-12 13:36:16 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 713/3300 [2:50:19<10:52:02, 15.12s/it]2026-04-12 13:36:31,708 | INFO | train_grpo_train | metrics_logged mode=train step=713

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 7.842424242424243e-06, 'num_tokens': 1443665.0, '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': 0.38649463653564453, 'rewards/meter/std': 0.35950174927711487, 'rewards/count_adherence/mean': 0.8999999761581421, 'rewards/count_adherence/std': 0.06172133609652519, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9993645548820496, 'rewards/lexical_plausibility/std': 0.0017972891218960285, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.001342153875157237, 'rewards/repeat_penalty/std': 0.0037961844354867935, 'rewards/near_duplicate_penalty/mean': 0.449945330619812, 'rewards/near_duplicate_penalty/std': 0.22313697636127472, 'rewards/opening_diversity/mean': 0.5614316463470459, 'rewards/opening_diversity/std': 0.35128137469291687, 'rewards/distinct_2/mean': 0.0018823768477886915, 'rewards/distinct_2/std': 0.005324165802448988, 'rewards/total_composite/mean': 0.03751775622367859, 'rewards/total_composite/std': 0.045408945530653, 'reward': 0.03751775622367859, 'reward_std': 0.0454089418053627, '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': 0.03751775622367859, 'reward_meter_mean': 0.38649463653564453, 'reward_meter_std': 0.35950174927711487, 'reward_count_adherence_mean': 0.8999999761581421, 'reward_count_adherence_std': 0.06172133609652519, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9993645548820496, 'reward_lexical_plausibility_std': 0.0017972891218960285, 'reward_repeat_penalty_mean': 0.001342153875157237, 'reward_repeat_penalty_std': 0.0037961844354867935, 'reward_near_duplicate_penalty_mean': 0.449945330619812, 'reward_near_duplicate_penalty_std': 0.22313697636127472, 'reward_opening_diversity_mean': 0.5614316463470459, 'reward_opening_diversity_std': 0.35128137469291687, 'reward_distinct_2_mean': 0.0018823768477886915, 'reward_distinct_2_std': 0.005324165802448988, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.03751775622367859, 'reward_total_composite_std': 0.045408945530653, 'epoch': 0.03}

 22%|██▏       | 713/3300 [2:50:19<10:52:02, 15.12s/it]INFO 04-12 13:36:31 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 714/3300 [2:50:30<9:59:22, 13.91s/it] 2026-04-12 13:36:42,780 | INFO | train_grpo_train | metrics_logged mode=train step=714

                                                      
{'loss': 0.0576, 'grad_norm': 4.135997772216797, 'learning_rate': 7.83939393939394e-06, 'num_tokens': 1447134.0, 'completions/mean_length': 249.625, 'completions/min_length': 213.0, 'completions/max_length': 267.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 249.625, 'completions/min_terminated_length': 213.0, 'completions/max_terminated_length': 267.0, 'rewards/meter/mean': 0.763771653175354, 'rewards/meter/std': 0.3426404595375061, 'rewards/count_adherence/mean': 0.9107142686843872, 'rewards/count_adherence/std': 0.10628911107778549, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9895833730697632, 'rewards/lexical_plausibility/std': 0.029462775215506554, 'rewards/judge_quality/mean': 0.13750000298023224, 'rewards/judge_quality/std': 0.0353553406894207, 'rewards/repeat_penalty/mean': 0.341755211353302, 'rewards/repeat_penalty/std': 0.2517761290073395, 'rewards/near_duplicate_penalty/mean': 0.7890658378601074, 'rewards/near_duplicate_penalty/std': 0.13595765829086304, 'rewards/opening_diversity/mean': 0.9317245483398438, 'rewards/opening_diversity/std': 0.12083348631858826, 'rewards/distinct_2/mean': 0.4419478178024292, 'rewards/distinct_2/std': 0.27744823694229126, 'rewards/total_composite/mean': 0.09026844799518585, 'rewards/total_composite/std': 0.061080846935510635, 'reward': 0.09026844799518585, 'reward_std': 0.06108083948493004, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08282623440027237, 'sampling/sampling_logp_difference/max': 1.605752944946289, 'sampling/importance_sampling_ratio/min': 0.2007383555173874, 'sampling/importance_sampling_ratio/mean': 1.0089975595474243, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.47546498849987984, 'clip_ratio/low_mean': 0.019779802300035954, 'clip_ratio/low_min': 0.019779802300035954, 'clip_ratio/high_mean': 0.05101823341101408, 'clip_ratio/high_max': 0.05101823341101408, 'clip_ratio/region_mean': 0.07079803571105003, 'reward_total_mean': 0.09026844799518585, 'reward_meter_mean': 0.763771653175354, 'reward_meter_std': 0.3426404595375061, 'reward_count_adherence_mean': 0.9107142686843872, 'reward_count_adherence_std': 0.10628911107778549, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9895833730697632, 'reward_lexical_plausibility_std': 0.029462775215506554, 'reward_repeat_penalty_mean': 0.341755211353302, 'reward_repeat_penalty_std': 0.2517761290073395, 'reward_near_duplicate_penalty_mean': 0.7890658378601074, 'reward_near_duplicate_penalty_std': 0.13595765829086304, 'reward_opening_diversity_mean': 0.9317245483398438, 'reward_opening_diversity_std': 0.12083348631858826, 'reward_distinct_2_mean': 0.4419478178024292, 'reward_distinct_2_std': 0.27744823694229126, 'reward_judge_quality_mean': 0.13750000298023224, 'reward_judge_quality_std': 0.0353553406894207, 'reward_total_composite_mean': 0.09026844799518585, 'reward_total_composite_std': 0.061080846935510635, 'epoch': 0.03}

 22%|██▏       | 714/3300 [2:50:30<9:59:22, 13.91s/it]INFO 04-12 13:36:43 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 715/3300 [2:50:44<9:58:41, 13.90s/it]2026-04-12 13:36:56,647 | INFO | train_grpo_train | metrics_logged mode=train step=715

                                                      
{'loss': -0.1372, 'grad_norm': 3.8975143432617188, 'learning_rate': 7.836363636363638e-06, 'num_tokens': 1449201.0, 'completions/mean_length': 211.375, 'completions/min_length': 87.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 111.16667175292969, 'completions/min_terminated_length': 87.0, 'completions/max_terminated_length': 126.0, 'rewards/meter/mean': 0.3653896450996399, 'rewards/meter/std': 0.3022627532482147, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.26726123690605164, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.953475296497345, 'rewards/lexical_plausibility/std': 0.08626838028430939, 'rewards/judge_quality/mean': 0.2875000238418579, 'rewards/judge_quality/std': 0.1505940705537796, 'rewards/repeat_penalty/mean': 0.88077712059021, 'rewards/repeat_penalty/std': 0.08017028123140335, 'rewards/near_duplicate_penalty/mean': 0.9692107439041138, 'rewards/near_duplicate_penalty/std': 0.0186021588742733, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9402357339859009, 'rewards/distinct_2/std': 0.05703572556376457, 'rewards/total_composite/mean': 0.12513285875320435, 'rewards/total_composite/std': 0.11413661390542984, 'reward': 0.12513285875320435, 'reward_std': 0.11413661390542984, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12337290495634079, 'sampling/sampling_logp_difference/max': 2.0883889198303223, 'sampling/importance_sampling_ratio/min': 0.12388657033443451, 'sampling/importance_sampling_ratio/mean': 1.0229735374450684, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6158076301217079, 'clip_ratio/low_mean': 0.026057031005620956, 'clip_ratio/low_min': 0.026057031005620956, 'clip_ratio/high_mean': 0.062277442775666714, 'clip_ratio/high_max': 0.062277442775666714, 'clip_ratio/region_mean': 0.08833447378128767, 'reward_total_mean': 0.12513285875320435, 'reward_meter_mean': 0.3653896450996399, 'reward_meter_std': 0.3022627532482147, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.26726123690605164, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.953475296497345, 'reward_lexical_plausibility_std': 0.08626838028430939, 'reward_repeat_penalty_mean': 0.88077712059021, 'reward_repeat_penalty_std': 0.08017028123140335, 'reward_near_duplicate_penalty_mean': 0.9692107439041138, 'reward_near_duplicate_penalty_std': 0.0186021588742733, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9402357339859009, 'reward_distinct_2_std': 0.05703572556376457, 'reward_judge_quality_mean': 0.2875000238418579, 'reward_judge_quality_std': 0.1505940705537796, 'reward_total_composite_mean': 0.12513285875320435, 'reward_total_composite_std': 0.11413661390542984, 'epoch': 0.03}

 22%|██▏       | 715/3300 [2:50:44<9:58:41, 13.90s/it]INFO 04-12 13:36:56 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 716/3300 [2:50:58<10:00:14, 13.94s/it]2026-04-12 13:37:10,688 | INFO | train_grpo_train | metrics_logged mode=train step=716

                                                       
{'loss': -0.0798, 'grad_norm': 4.312860488891602, 'learning_rate': 7.833333333333333e-06, 'num_tokens': 1451591.0, 'completions/mean_length': 189.75, 'completions/min_length': 126.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 143.71429443359375, 'completions/min_terminated_length': 126.0, 'completions/max_terminated_length': 161.0, 'rewards/meter/mean': 0.5335475206375122, 'rewards/meter/std': 0.4025692343711853, 'rewards/count_adherence/mean': 0.84375, 'rewards/count_adherence/std': 0.35197150707244873, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9378854036331177, 'rewards/lexical_plausibility/std': 0.17568671703338623, 'rewards/judge_quality/mean': 0.1875, 'rewards/judge_quality/std': 0.1060660183429718, 'rewards/repeat_penalty/mean': 0.4815531373023987, 'rewards/repeat_penalty/std': 0.27292245626449585, 'rewards/near_duplicate_penalty/mean': 0.7996932864189148, 'rewards/near_duplicate_penalty/std': 0.11553988605737686, 'rewards/opening_diversity/mean': 0.9609375, 'rewards/opening_diversity/std': 0.08799287676811218, 'rewards/distinct_2/mean': 0.6112323999404907, 'rewards/distinct_2/std': 0.2175302803516388, 'rewards/total_composite/mean': 0.11849026381969452, 'rewards/total_composite/std': 0.11608956009149551, 'reward': 0.11849026381969452, 'reward_std': 0.11608955264091492, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1107170507311821, 'sampling/sampling_logp_difference/max': 3.2527613639831543, 'sampling/importance_sampling_ratio/min': 0.038667287677526474, 'sampling/importance_sampling_ratio/mean': 1.0102763175964355, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5769151635468006, 'clip_ratio/low_mean': 0.033641617745161057, 'clip_ratio/low_min': 0.033641617745161057, 'clip_ratio/high_mean': 0.054977825842797756, 'clip_ratio/high_max': 0.054977825842797756, 'clip_ratio/region_mean': 0.08861944358795881, 'reward_total_mean': 0.11849026381969452, 'reward_meter_mean': 0.5335475206375122, 'reward_meter_std': 0.4025692343711853, 'reward_count_adherence_mean': 0.84375, 'reward_count_adherence_std': 0.35197150707244873, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9378854036331177, 'reward_lexical_plausibility_std': 0.17568671703338623, 'reward_repeat_penalty_mean': 0.4815531373023987, 'reward_repeat_penalty_std': 0.27292245626449585, 'reward_near_duplicate_penalty_mean': 0.7996932864189148, 'reward_near_duplicate_penalty_std': 0.11553988605737686, 'reward_opening_diversity_mean': 0.9609375, 'reward_opening_diversity_std': 0.08799287676811218, 'reward_distinct_2_mean': 0.6112323999404907, 'reward_distinct_2_std': 0.2175302803516388, 'reward_judge_quality_mean': 0.1875, 'reward_judge_quality_std': 0.1060660183429718, 'reward_total_composite_mean': 0.11849026381969452, 'reward_total_composite_std': 0.11608956009149551, 'epoch': 0.03}

 22%|██▏       | 716/3300 [2:50:58<10:00:14, 13.94s/it]INFO 04-12 13:37:10 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 717/3300 [2:51:08<9:13:25, 12.86s/it] 2026-04-12 13:37:21,014 | INFO | train_grpo_train | metrics_logged mode=train step=717

                                                      
{'loss': 0.0681, 'grad_norm': 8.300482749938965, 'learning_rate': 7.83030303030303e-06, 'num_tokens': 1454461.0, 'completions/mean_length': 151.75, 'completions/min_length': 120.0, 'completions/max_length': 184.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 151.75, 'completions/min_terminated_length': 120.0, 'completions/max_terminated_length': 184.0, 'rewards/meter/mean': 0.5869561433792114, 'rewards/meter/std': 0.4143819510936737, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.1035098284482956, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.984375, 'rewards/lexical_plausibility/std': 0.04419417306780815, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.1060660183429718, 'rewards/repeat_penalty/mean': 0.3283223807811737, 'rewards/repeat_penalty/std': 0.28635236620903015, 'rewards/near_duplicate_penalty/mean': 0.7320524454116821, 'rewards/near_duplicate_penalty/std': 0.1146959513425827, 'rewards/opening_diversity/mean': 0.8026041984558105, 'rewards/opening_diversity/std': 0.33866962790489197, 'rewards/distinct_2/mean': 0.5139428377151489, 'rewards/distinct_2/std': 0.2744678556919098, 'rewards/total_composite/mean': 0.1502363085746765, 'rewards/total_composite/std': 0.12501950562000275, 'reward': 0.1502363085746765, 'reward_std': 0.12501949071884155, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09794704616069794, 'sampling/sampling_logp_difference/max': 1.7185297012329102, 'sampling/importance_sampling_ratio/min': 0.17932963371276855, 'sampling/importance_sampling_ratio/mean': 1.00253164768219, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5102434642612934, 'clip_ratio/low_mean': 0.04951613815501332, 'clip_ratio/low_min': 0.04951613815501332, 'clip_ratio/high_mean': 0.044275776483118534, 'clip_ratio/high_max': 0.044275776483118534, 'clip_ratio/region_mean': 0.09379191463813186, 'reward_total_mean': 0.1502363085746765, 'reward_meter_mean': 0.5869561433792114, 'reward_meter_std': 0.4143819510936737, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.1035098284482956, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.984375, 'reward_lexical_plausibility_std': 0.04419417306780815, 'reward_repeat_penalty_mean': 0.3283223807811737, 'reward_repeat_penalty_std': 0.28635236620903015, 'reward_near_duplicate_penalty_mean': 0.7320524454116821, 'reward_near_duplicate_penalty_std': 0.1146959513425827, 'reward_opening_diversity_mean': 0.8026041984558105, 'reward_opening_diversity_std': 0.33866962790489197, 'reward_distinct_2_mean': 0.5139428377151489, 'reward_distinct_2_std': 0.2744678556919098, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.1060660183429718, 'reward_total_composite_mean': 0.1502363085746765, 'reward_total_composite_std': 0.12501950562000275, 'epoch': 0.03}

 22%|██▏       | 717/3300 [2:51:08<9:13:25, 12.86s/it]INFO 04-12 13:37:21 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 718/3300 [2:51:22<9:22:22, 13.07s/it]2026-04-12 13:37:34,578 | INFO | train_grpo_train | metrics_logged mode=train step=718

                                                      
{'loss': -0.0676, 'grad_norm': 3.2426297664642334, 'learning_rate': 7.827272727272728e-06, 'num_tokens': 1456035.0, 'completions/mean_length': 167.75, 'completions/min_length': 46.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 53.0, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.5163996815681458, 'rewards/meter/std': 0.45872992277145386, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9077537655830383, 'rewards/lexical_plausibility/std': 0.1766103208065033, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.1767766922712326, 'rewards/repeat_penalty/mean': 0.9234302043914795, 'rewards/repeat_penalty/std': 0.08610820025205612, 'rewards/near_duplicate_penalty/mean': 0.9635096788406372, 'rewards/near_duplicate_penalty/std': 0.03608851507306099, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.990384578704834, 'rewards/distinct_2/std': 0.027196412906050682, 'rewards/total_composite/mean': 0.19901247322559357, 'rewards/total_composite/std': 0.19665133953094482, 'reward': 0.19901247322559357, 'reward_std': 0.19665133953094482, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14281298220157623, 'sampling/sampling_logp_difference/max': 1.3788189888000488, 'sampling/importance_sampling_ratio/min': 0.2518758475780487, 'sampling/importance_sampling_ratio/mean': 1.0090091228485107, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7193535938858986, 'clip_ratio/low_mean': 0.020877009257674217, 'clip_ratio/low_min': 0.020877009257674217, 'clip_ratio/high_mean': 0.08075364772230387, 'clip_ratio/high_max': 0.08075364772230387, 'clip_ratio/region_mean': 0.10163065697997808, 'reward_total_mean': 0.19901247322559357, 'reward_meter_mean': 0.5163996815681458, 'reward_meter_std': 0.45872992277145386, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9077537655830383, 'reward_lexical_plausibility_std': 0.1766103208065033, 'reward_repeat_penalty_mean': 0.9234302043914795, 'reward_repeat_penalty_std': 0.08610820025205612, 'reward_near_duplicate_penalty_mean': 0.9635096788406372, 'reward_near_duplicate_penalty_std': 0.03608851507306099, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.990384578704834, 'reward_distinct_2_std': 0.027196412906050682, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.1767766922712326, 'reward_total_composite_mean': 0.19901247322559357, 'reward_total_composite_std': 0.19665133953094482, 'epoch': 0.03}

 22%|██▏       | 718/3300 [2:51:22<9:22:22, 13.07s/it]INFO 04-12 13:37:34 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 719/3300 [2:51:32<8:37:20, 12.03s/it]2026-04-12 13:37:44,178 | INFO | train_grpo_train | metrics_logged mode=train step=719

                                                      
{'loss': 0.0118, 'grad_norm': 7.891262054443359, 'learning_rate': 7.824242424242425e-06, 'num_tokens': 1458436.0, 'completions/mean_length': 133.125, 'completions/min_length': 119.0, 'completions/max_length': 145.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 133.125, 'completions/min_terminated_length': 119.0, 'completions/max_terminated_length': 145.0, 'rewards/meter/mean': 0.810893714427948, 'rewards/meter/std': 0.27897176146507263, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.20000000298023224, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.6247476935386658, 'rewards/repeat_penalty/std': 0.1748056262731552, 'rewards/near_duplicate_penalty/mean': 0.8178353309631348, 'rewards/near_duplicate_penalty/std': 0.0779615268111229, 'rewards/opening_diversity/mean': 0.984375, 'rewards/opening_diversity/std': 0.0289318785071373, 'rewards/distinct_2/mean': 0.7636566162109375, 'rewards/distinct_2/std': 0.148441344499588, 'rewards/total_composite/mean': 0.16775871813297272, 'rewards/total_composite/std': 0.10743365436792374, 'reward': 0.16775871813297272, 'reward_std': 0.10743366181850433, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11007557809352875, 'sampling/sampling_logp_difference/max': 2.2130088806152344, 'sampling/importance_sampling_ratio/min': 0.10937107354402542, 'sampling/importance_sampling_ratio/mean': 0.9958500862121582, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7274518944323063, 'clip_ratio/low_mean': 0.06906360620632768, 'clip_ratio/low_min': 0.06906360620632768, 'clip_ratio/high_mean': 0.02703285962343216, 'clip_ratio/high_max': 0.02703285962343216, 'clip_ratio/region_mean': 0.09609646582975984, 'reward_total_mean': 0.16775871813297272, 'reward_meter_mean': 0.810893714427948, 'reward_meter_std': 0.27897176146507263, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6247476935386658, 'reward_repeat_penalty_std': 0.1748056262731552, 'reward_near_duplicate_penalty_mean': 0.8178353309631348, 'reward_near_duplicate_penalty_std': 0.0779615268111229, 'reward_opening_diversity_mean': 0.984375, 'reward_opening_diversity_std': 0.0289318785071373, 'reward_distinct_2_mean': 0.7636566162109375, 'reward_distinct_2_std': 0.148441344499588, 'reward_judge_quality_mean': 0.20000000298023224, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.16775871813297272, 'reward_total_composite_std': 0.10743365436792374, 'epoch': 0.03}

 22%|██▏       | 719/3300 [2:51:32<8:37:20, 12.03s/it]INFO 04-12 13:37:44 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 720/3300 [2:51:43<8:23:51, 11.72s/it]2026-04-12 13:37:55,172 | INFO | train_grpo_train | metrics_logged mode=train step=720

                                                      
{'loss': 0.0449, 'grad_norm': 5.318984031677246, 'learning_rate': 7.821212121212122e-06, 'num_tokens': 1461868.0, 'completions/mean_length': 232.0, 'completions/min_length': 204.0, 'completions/max_length': 247.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 232.0, 'completions/min_terminated_length': 204.0, 'completions/max_terminated_length': 247.0, 'rewards/meter/mean': 0.3841937184333801, 'rewards/meter/std': 0.3249920904636383, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.17500001192092896, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/repeat_penalty/mean': 0.47536277770996094, 'rewards/repeat_penalty/std': 0.30891862511634827, 'rewards/near_duplicate_penalty/mean': 0.8349100947380066, 'rewards/near_duplicate_penalty/std': 0.09276093542575836, 'rewards/opening_diversity/mean': 0.9642857313156128, 'rewards/opening_diversity/std': 0.0874817818403244, 'rewards/distinct_2/mean': 0.5841090679168701, 'rewards/distinct_2/std': 0.2916153371334076, 'rewards/total_composite/mean': 0.06952423602342606, 'rewards/total_composite/std': 0.0673690140247345, 'reward': 0.06952423602342606, 'reward_std': 0.0673690140247345, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09519623965024948, 'sampling/sampling_logp_difference/max': 1.6202712059020996, 'sampling/importance_sampling_ratio/min': 0.19784502685070038, 'sampling/importance_sampling_ratio/mean': 1.0090858936309814, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5764502733945847, 'clip_ratio/low_mean': 0.041640359442681074, 'clip_ratio/low_min': 0.041640359442681074, 'clip_ratio/high_mean': 0.05514508578926325, 'clip_ratio/high_max': 0.05514508578926325, 'clip_ratio/region_mean': 0.09678544523194432, 'reward_total_mean': 0.06952423602342606, 'reward_meter_mean': 0.3841937184333801, 'reward_meter_std': 0.3249920904636383, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.47536277770996094, 'reward_repeat_penalty_std': 0.30891862511634827, 'reward_near_duplicate_penalty_mean': 0.8349100947380066, 'reward_near_duplicate_penalty_std': 0.09276093542575836, 'reward_opening_diversity_mean': 0.9642857313156128, 'reward_opening_diversity_std': 0.0874817818403244, 'reward_distinct_2_mean': 0.5841090679168701, 'reward_distinct_2_std': 0.2916153371334076, 'reward_judge_quality_mean': 0.17500001192092896, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.06952423602342606, 'reward_total_composite_std': 0.0673690140247345, 'epoch': 0.03}

 22%|██▏       | 720/3300 [2:51:43<8:23:51, 11.72s/it]INFO 04-12 13:37:55 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 721/3300 [2:51:51<7:41:46, 10.74s/it]2026-04-12 13:38:03,642 | INFO | train_grpo_train | metrics_logged mode=train step=721

                                                      
{'loss': 0.025, 'grad_norm': 10.545655250549316, 'learning_rate': 7.81818181818182e-06, 'num_tokens': 1463683.0, 'completions/mean_length': 60.875, 'completions/min_length': 53.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.875, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.7616569995880127, 'rewards/meter/std': 0.31978484988212585, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.35874998569488525, 'rewards/judge_quality/std': 0.2547512948513031, 'rewards/repeat_penalty/mean': 0.789879322052002, 'rewards/repeat_penalty/std': 0.1283789724111557, 'rewards/near_duplicate_penalty/mean': 0.8887174725532532, 'rewards/near_duplicate_penalty/std': 0.07181558012962341, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8841814398765564, 'rewards/distinct_2/std': 0.07543253898620605, 'rewards/total_composite/mean': 0.29626649618148804, 'rewards/total_composite/std': 0.28620389103889465, 'reward': 0.29626649618148804, 'reward_std': 0.28620386123657227, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1434631496667862, 'sampling/sampling_logp_difference/max': 1.7655353546142578, 'sampling/importance_sampling_ratio/min': 0.17109516263008118, 'sampling/importance_sampling_ratio/mean': 1.0185937881469727, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0090814530849457, 'clip_ratio/low_mean': 0.0695754736661911, 'clip_ratio/low_min': 0.0695754736661911, 'clip_ratio/high_mean': 0.06644419766962528, 'clip_ratio/high_max': 0.06644419766962528, 'clip_ratio/region_mean': 0.13601967133581638, 'reward_total_mean': 0.29626649618148804, 'reward_meter_mean': 0.7616569995880127, 'reward_meter_std': 0.31978484988212585, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.789879322052002, 'reward_repeat_penalty_std': 0.1283789724111557, 'reward_near_duplicate_penalty_mean': 0.8887174725532532, 'reward_near_duplicate_penalty_std': 0.07181558012962341, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8841814398765564, 'reward_distinct_2_std': 0.07543253898620605, 'reward_judge_quality_mean': 0.35874998569488525, 'reward_judge_quality_std': 0.2547512948513031, 'reward_total_composite_mean': 0.29626649618148804, 'reward_total_composite_std': 0.28620389103889465, 'epoch': 0.03}

 22%|██▏       | 721/3300 [2:51:51<7:41:46, 10.74s/it]INFO 04-12 13:38:03 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 722/3300 [2:52:05<8:16:37, 11.56s/it]2026-04-12 13:38:17,101 | INFO | train_grpo_train | metrics_logged mode=train step=722

                                                      
{'loss': 0.0401, 'grad_norm': 5.4081501960754395, 'learning_rate': 7.815151515151515e-06, 'num_tokens': 1465108.0, 'completions/mean_length': 94.125, 'completions/min_length': 21.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 34.42857360839844, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.5046600103378296, 'rewards/meter/std': 0.4424968659877777, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.4629100561141968, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.951648473739624, 'rewards/lexical_plausibility/std': 0.13675887882709503, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.35722893476486206, 'rewards/repeat_penalty/mean': 0.7705550789833069, 'rewards/repeat_penalty/std': 0.09742417931556702, 'rewards/near_duplicate_penalty/mean': 0.9825732111930847, 'rewards/near_duplicate_penalty/std': 0.04929042607545853, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9715099334716797, 'rewards/distinct_2/std': 0.08058197051286697, 'rewards/total_composite/mean': 0.28435760736465454, 'rewards/total_composite/std': 0.3203091323375702, 'reward': 0.28435760736465454, 'reward_std': 0.3203091323375702, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1381158083677292, 'sampling/sampling_logp_difference/max': 1.1411648988723755, 'sampling/importance_sampling_ratio/min': 0.4382007122039795, 'sampling/importance_sampling_ratio/mean': 1.0286113023757935, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8700064122676849, 'clip_ratio/low_mean': 0.03146777022629976, 'clip_ratio/low_min': 0.03146777022629976, 'clip_ratio/high_mean': 0.08177659101784229, 'clip_ratio/high_max': 0.08177659101784229, 'clip_ratio/region_mean': 0.11324436124414206, 'reward_total_mean': 0.28435760736465454, 'reward_meter_mean': 0.5046600103378296, 'reward_meter_std': 0.4424968659877777, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.4629100561141968, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.951648473739624, 'reward_lexical_plausibility_std': 0.13675887882709503, 'reward_repeat_penalty_mean': 0.7705550789833069, 'reward_repeat_penalty_std': 0.09742417931556702, 'reward_near_duplicate_penalty_mean': 0.9825732111930847, 'reward_near_duplicate_penalty_std': 0.04929042607545853, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9715099334716797, 'reward_distinct_2_std': 0.08058197051286697, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.35722893476486206, 'reward_total_composite_mean': 0.28435760736465454, 'reward_total_composite_std': 0.3203091323375702, 'epoch': 0.03}

 22%|██▏       | 722/3300 [2:52:05<8:16:37, 11.56s/it]INFO 04-12 13:38:17 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 723/3300 [2:52:18<8:44:40, 12.22s/it]2026-04-12 13:38:30,859 | INFO | train_grpo_train | metrics_logged mode=train step=723

                                                      
{'loss': 0.102, 'grad_norm': 0.5549939870834351, 'learning_rate': 7.812121212121213e-06, 'num_tokens': 1467306.0, 'completions/mean_length': 508.75, 'completions/min_length': 486.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.875, 'completions/mean_terminated_length': 486.0, 'completions/min_terminated_length': 486.0, 'completions/max_terminated_length': 486.0, 'rewards/meter/mean': 0.8639713525772095, 'rewards/meter/std': 0.15796209871768951, 'rewards/count_adherence/mean': 0.8482142686843872, 'rewards/count_adherence/std': 0.3181522488594055, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.11249999701976776, 'rewards/judge_quality/std': 0.051754921674728394, 'rewards/repeat_penalty/mean': 0.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 0.4627264142036438, 'rewards/near_duplicate_penalty/std': 0.22879651188850403, 'rewards/opening_diversity/mean': 0.8230406641960144, 'rewards/opening_diversity/std': 0.08088110387325287, 'rewards/distinct_2/mean': 0.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.0841808170080185, 'rewards/total_composite/std': 0.04921037703752518, 'reward': 0.0841808170080185, 'reward_std': 0.04921037331223488, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.013573627918958664, 'sampling/sampling_logp_difference/max': 0.5130642652511597, 'sampling/importance_sampling_ratio/min': 0.5986583232879639, 'sampling/importance_sampling_ratio/mean': 1.0011045932769775, 'sampling/importance_sampling_ratio/max': 1.5862984657287598, 'entropy': 0.009813507087528706, 'clip_ratio/low_mean': 0.0015432098880410194, 'clip_ratio/low_min': 0.0015432098880410194, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0015432098880410194, 'reward_total_mean': 0.0841808170080185, 'reward_meter_mean': 0.8639713525772095, 'reward_meter_std': 0.15796209871768951, 'reward_count_adherence_mean': 0.8482142686843872, 'reward_count_adherence_std': 0.3181522488594055, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.0, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 0.4627264142036438, 'reward_near_duplicate_penalty_std': 0.22879651188850403, 'reward_opening_diversity_mean': 0.8230406641960144, 'reward_opening_diversity_std': 0.08088110387325287, 'reward_distinct_2_mean': 0.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.11249999701976776, 'reward_judge_quality_std': 0.051754921674728394, 'reward_total_composite_mean': 0.0841808170080185, 'reward_total_composite_std': 0.04921037703752518, 'epoch': 0.03}

 22%|██▏       | 723/3300 [2:52:18<8:44:40, 12.22s/it]INFO 04-12 13:38:31 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 724/3300 [2:52:30<8:32:16, 11.93s/it]2026-04-12 13:38:42,122 | INFO | train_grpo_train | metrics_logged mode=train step=724

                                                      
{'loss': -0.0415, 'grad_norm': 3.912525177001953, 'learning_rate': 7.80909090909091e-06, 'num_tokens': 1471331.0, 'completions/mean_length': 285.125, 'completions/min_length': 256.0, 'completions/max_length': 309.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 285.125, 'completions/min_terminated_length': 256.0, 'completions/max_terminated_length': 309.0, 'rewards/meter/mean': 0.9829435348510742, 'rewards/meter/std': 0.03629862517118454, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.06681530922651291, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.06332159042358398, 'rewards/repeat_penalty/std': 0.046767961233854294, 'rewards/near_duplicate_penalty/mean': 0.7141451835632324, 'rewards/near_duplicate_penalty/std': 0.020828045904636383, 'rewards/opening_diversity/mean': 0.8824776411056519, 'rewards/opening_diversity/std': 0.11834309995174408, 'rewards/distinct_2/mean': 0.15145601332187653, 'rewards/distinct_2/std': 0.10429301112890244, 'rewards/total_composite/mean': 0.1381019502878189, 'rewards/total_composite/std': 0.00931702833622694, 'reward': 0.1381019502878189, 'reward_std': 0.009317024610936642, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06171497702598572, 'sampling/sampling_logp_difference/max': 1.7601337432861328, 'sampling/importance_sampling_ratio/min': 0.17202186584472656, 'sampling/importance_sampling_ratio/mean': 1.002913475036621, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.34784314036369324, 'clip_ratio/low_mean': 0.03085204283706844, 'clip_ratio/low_min': 0.03085204283706844, 'clip_ratio/high_mean': 0.024414620827883482, 'clip_ratio/high_max': 0.024414620827883482, 'clip_ratio/region_mean': 0.05526666366495192, 'reward_total_mean': 0.1381019502878189, 'reward_meter_mean': 0.9829435348510742, 'reward_meter_std': 0.03629862517118454, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.06681530922651291, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.06332159042358398, 'reward_repeat_penalty_std': 0.046767961233854294, 'reward_near_duplicate_penalty_mean': 0.7141451835632324, 'reward_near_duplicate_penalty_std': 0.020828045904636383, 'reward_opening_diversity_mean': 0.8824776411056519, 'reward_opening_diversity_std': 0.11834309995174408, 'reward_distinct_2_mean': 0.15145601332187653, 'reward_distinct_2_std': 0.10429301112890244, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.1381019502878189, 'reward_total_composite_std': 0.00931702833622694, 'epoch': 0.03}

 22%|██▏       | 724/3300 [2:52:30<8:32:16, 11.93s/it]INFO 04-12 13:38:42 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 725/3300 [2:52:38<7:47:43, 10.90s/it]2026-04-12 13:38:50,612 | INFO | train_grpo_train | metrics_logged mode=train step=725

                                                      
{'loss': 0.0044, 'grad_norm': 13.008666038513184, 'learning_rate': 7.806060606060607e-06, 'num_tokens': 1473120.0, 'completions/mean_length': 56.625, 'completions/min_length': 42.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.625, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.7698050141334534, 'rewards/meter/std': 0.18485382199287415, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.45500001311302185, 'rewards/judge_quality/std': 0.31883716583251953, 'rewards/repeat_penalty/mean': 0.7339037656784058, 'rewards/repeat_penalty/std': 0.4061550199985504, 'rewards/near_duplicate_penalty/mean': 0.8356177806854248, 'rewards/near_duplicate_penalty/std': 0.22047339379787445, 'rewards/opening_diversity/mean': 0.890625, 'rewards/opening_diversity/std': 0.2259652018547058, 'rewards/distinct_2/mean': 0.8368016481399536, 'rewards/distinct_2/std': 0.30542340874671936, 'rewards/total_composite/mean': 0.356789767742157, 'rewards/total_composite/std': 0.2745116651058197, 'reward': 0.356789767742157, 'reward_std': 0.2745116651058197, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13698266446590424, 'sampling/sampling_logp_difference/max': 1.2988669872283936, 'sampling/importance_sampling_ratio/min': 0.2728407382965088, 'sampling/importance_sampling_ratio/mean': 1.0128178596496582, 'sampling/importance_sampling_ratio/max': 1.9994455575942993, 'entropy': 0.9383715838193893, 'clip_ratio/low_mean': 0.05205825716257095, 'clip_ratio/low_min': 0.05205825716257095, 'clip_ratio/high_mean': 0.07010691985487938, 'clip_ratio/high_max': 0.07010691985487938, 'clip_ratio/region_mean': 0.12216517701745033, 'reward_total_mean': 0.356789767742157, 'reward_meter_mean': 0.7698050141334534, 'reward_meter_std': 0.18485382199287415, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7339037656784058, 'reward_repeat_penalty_std': 0.4061550199985504, 'reward_near_duplicate_penalty_mean': 0.8356177806854248, 'reward_near_duplicate_penalty_std': 0.22047339379787445, 'reward_opening_diversity_mean': 0.890625, 'reward_opening_diversity_std': 0.2259652018547058, 'reward_distinct_2_mean': 0.8368016481399536, 'reward_distinct_2_std': 0.30542340874671936, 'reward_judge_quality_mean': 0.45500001311302185, 'reward_judge_quality_std': 0.31883716583251953, 'reward_total_composite_mean': 0.356789767742157, 'reward_total_composite_std': 0.2745116651058197, 'epoch': 0.03}

 22%|██▏       | 725/3300 [2:52:38<7:47:43, 10.90s/it]INFO 04-12 13:38:50 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 726/3300 [2:52:57<9:29:02, 13.26s/it]2026-04-12 13:39:09,396 | INFO | train_grpo_train | metrics_logged mode=train step=726

                                                      
{'loss': -0.1141, 'grad_norm': 5.9563727378845215, 'learning_rate': 7.803030303030303e-06, 'num_tokens': 1475176.0, 'completions/mean_length': 158.0, 'completions/min_length': 97.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 107.42857360839844, 'completions/min_terminated_length': 97.0, 'completions/max_terminated_length': 118.0, 'rewards/meter/mean': 0.7789549827575684, 'rewards/meter/std': 0.2228599637746811, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.2357022762298584, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9476962089538574, 'rewards/lexical_plausibility/std': 0.14793740212917328, 'rewards/judge_quality/mean': 0.23750001192092896, 'rewards/judge_quality/std': 0.12464234977960587, 'rewards/repeat_penalty/mean': 0.6963146924972534, 'rewards/repeat_penalty/std': 0.2755965292453766, 'rewards/near_duplicate_penalty/mean': 0.9046881794929504, 'rewards/near_duplicate_penalty/std': 0.0705389678478241, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.8293020725250244, 'rewards/distinct_2/std': 0.2154710590839386, 'rewards/total_composite/mean': 0.16904276609420776, 'rewards/total_composite/std': 0.10946167260408401, 'reward': 0.16904276609420776, 'reward_std': 0.10946167260408401, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12301559001207352, 'sampling/sampling_logp_difference/max': 1.5293889045715332, 'sampling/importance_sampling_ratio/min': 0.216668039560318, 'sampling/importance_sampling_ratio/mean': 1.0064878463745117, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7016413509845734, 'clip_ratio/low_mean': 0.06256412249058485, 'clip_ratio/low_min': 0.06256412249058485, 'clip_ratio/high_mean': 0.059111873619258404, 'clip_ratio/high_max': 0.059111873619258404, 'clip_ratio/region_mean': 0.12167599610984325, 'reward_total_mean': 0.16904276609420776, 'reward_meter_mean': 0.7789549827575684, 'reward_meter_std': 0.2228599637746811, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.2357022762298584, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9476962089538574, 'reward_lexical_plausibility_std': 0.14793740212917328, 'reward_repeat_penalty_mean': 0.6963146924972534, 'reward_repeat_penalty_std': 0.2755965292453766, 'reward_near_duplicate_penalty_mean': 0.9046881794929504, 'reward_near_duplicate_penalty_std': 0.0705389678478241, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.8293020725250244, 'reward_distinct_2_std': 0.2154710590839386, 'reward_judge_quality_mean': 0.23750001192092896, 'reward_judge_quality_std': 0.12464234977960587, 'reward_total_composite_mean': 0.16904276609420776, 'reward_total_composite_std': 0.10946167260408401, 'epoch': 0.03}

 22%|██▏       | 726/3300 [2:52:57<9:29:02, 13.26s/it]INFO 04-12 13:39:09 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 727/3300 [2:53:06<8:33:18, 11.97s/it]2026-04-12 13:39:18,346 | INFO | train_grpo_train | metrics_logged mode=train step=727

                                                      
{'loss': 0.0469, 'grad_norm': 9.287501335144043, 'learning_rate': 7.800000000000002e-06, 'num_tokens': 1476992.0, 'completions/mean_length': 87.0, 'completions/min_length': 80.0, 'completions/max_length': 92.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 87.0, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 92.0, 'rewards/meter/mean': 0.8757219910621643, 'rewards/meter/std': 0.18546900153160095, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.591644287109375, 'rewards/repeat_penalty/std': 0.2018372267484665, 'rewards/near_duplicate_penalty/mean': 0.851108729839325, 'rewards/near_duplicate_penalty/std': 0.07145041227340698, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.6957238912582397, 'rewards/distinct_2/std': 0.17266994714736938, 'rewards/total_composite/mean': 0.321175754070282, 'rewards/total_composite/std': 0.08148026466369629, 'reward': 0.321175754070282, 'reward_std': 0.08148025721311569, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10866335779428482, 'sampling/sampling_logp_difference/max': 1.2884516716003418, 'sampling/importance_sampling_ratio/min': 0.2756973206996918, 'sampling/importance_sampling_ratio/mean': 1.0034369230270386, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6825176067650318, 'clip_ratio/low_mean': 0.035490380600094795, 'clip_ratio/low_min': 0.035490380600094795, 'clip_ratio/high_mean': 0.054789694491773844, 'clip_ratio/high_max': 0.054789694491773844, 'clip_ratio/region_mean': 0.09028007509186864, 'reward_total_mean': 0.321175754070282, 'reward_meter_mean': 0.8757219910621643, 'reward_meter_std': 0.18546900153160095, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.591644287109375, 'reward_repeat_penalty_std': 0.2018372267484665, 'reward_near_duplicate_penalty_mean': 0.851108729839325, 'reward_near_duplicate_penalty_std': 0.07145041227340698, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.6957238912582397, 'reward_distinct_2_std': 0.17266994714736938, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.321175754070282, 'reward_total_composite_std': 0.08148026466369629, 'epoch': 0.03}

 22%|██▏       | 727/3300 [2:53:06<8:33:18, 11.97s/it]INFO 04-12 13:39:18 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 728/3300 [2:53:18<8:40:18, 12.14s/it]2026-04-12 13:39:30,874 | INFO | train_grpo_train | metrics_logged mode=train step=728

                                                      
{'loss': 0.0458, 'grad_norm': 2.824350357055664, 'learning_rate': 7.796969696969697e-06, 'num_tokens': 1481540.0, 'completions/mean_length': 351.5, 'completions/min_length': 289.0, 'completions/max_length': 387.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 351.5, 'completions/min_terminated_length': 289.0, 'completions/max_terminated_length': 387.0, 'rewards/meter/mean': 0.9427692294120789, 'rewards/meter/std': 0.10438236594200134, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 0.5373981595039368, 'rewards/near_duplicate_penalty/std': 0.15793682634830475, 'rewards/opening_diversity/mean': 0.7524038553237915, 'rewards/opening_diversity/std': 0.26054221391677856, 'rewards/distinct_2/mean': 0.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.12963077425956726, 'rewards/total_composite/std': 0.014352576807141304, 'reward': 0.12963077425956726, 'reward_std': 0.014352577738463879, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.042539581656455994, 'sampling/sampling_logp_difference/max': 2.898378372192383, 'sampling/importance_sampling_ratio/min': 0.05511251837015152, 'sampling/importance_sampling_ratio/mean': 1.002425193786621, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.24230213835835457, 'clip_ratio/low_mean': 0.003552971640601754, 'clip_ratio/low_min': 0.003552971640601754, 'clip_ratio/high_mean': 0.03891419293358922, 'clip_ratio/high_max': 0.03891419293358922, 'clip_ratio/region_mean': 0.042467164574190974, 'reward_total_mean': 0.12963077425956726, 'reward_meter_mean': 0.9427692294120789, 'reward_meter_std': 0.10438236594200134, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.0, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 0.5373981595039368, 'reward_near_duplicate_penalty_std': 0.15793682634830475, 'reward_opening_diversity_mean': 0.7524038553237915, 'reward_opening_diversity_std': 0.26054221391677856, 'reward_distinct_2_mean': 0.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.12963077425956726, 'reward_total_composite_std': 0.014352576807141304, 'epoch': 0.03}

 22%|██▏       | 728/3300 [2:53:18<8:40:18, 12.14s/it]INFO 04-12 13:39:31 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 729/3300 [2:53:28<8:02:37, 11.26s/it]2026-04-12 13:39:40,100 | INFO | train_grpo_train | metrics_logged mode=train step=729

                                                      
{'loss': 0.0559, 'grad_norm': 9.724015235900879, 'learning_rate': 7.793939393939394e-06, 'num_tokens': 1483306.0, 'completions/mean_length': 64.75, 'completions/min_length': 55.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.75, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.6329939961433411, 'rewards/meter/std': 0.40112221240997314, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9887152910232544, 'rewards/lexical_plausibility/std': 0.022863244637846947, 'rewards/judge_quality/mean': 0.42499998211860657, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.9505146145820618, 'rewards/repeat_penalty/std': 0.05243025720119476, 'rewards/near_duplicate_penalty/mean': 0.958681046962738, 'rewards/near_duplicate_penalty/std': 0.031088965013623238, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9908424615859985, 'rewards/distinct_2/std': 0.025901345536112785, 'rewards/total_composite/mean': 0.2747950553894043, 'rewards/total_composite/std': 0.20810966193675995, 'reward': 0.2747950553894043, 'reward_std': 0.20810966193675995, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12471278756856918, 'sampling/sampling_logp_difference/max': 1.5095486640930176, 'sampling/importance_sampling_ratio/min': 0.22100970149040222, 'sampling/importance_sampling_ratio/mean': 0.9938512444496155, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7396027520298958, 'clip_ratio/low_mean': 0.06559466850012541, 'clip_ratio/low_min': 0.06559466850012541, 'clip_ratio/high_mean': 0.07234858348965645, 'clip_ratio/high_max': 0.07234858348965645, 'clip_ratio/region_mean': 0.13794325198978186, 'reward_total_mean': 0.2747950553894043, 'reward_meter_mean': 0.6329939961433411, 'reward_meter_std': 0.40112221240997314, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9887152910232544, 'reward_lexical_plausibility_std': 0.022863244637846947, 'reward_repeat_penalty_mean': 0.9505146145820618, 'reward_repeat_penalty_std': 0.05243025720119476, 'reward_near_duplicate_penalty_mean': 0.958681046962738, 'reward_near_duplicate_penalty_std': 0.031088965013623238, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9908424615859985, 'reward_distinct_2_std': 0.025901345536112785, 'reward_judge_quality_mean': 0.42499998211860657, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.2747950553894043, 'reward_total_composite_std': 0.20810966193675995, 'epoch': 0.03}

 22%|██▏       | 729/3300 [2:53:28<8:02:37, 11.26s/it]INFO 04-12 13:39:40 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 730/3300 [2:53:41<8:36:51, 12.07s/it]2026-04-12 13:39:54,035 | INFO | train_grpo_train | metrics_logged mode=train step=730

                                                      
{'loss': -0.0786, 'grad_norm': 5.605185508728027, 'learning_rate': 7.790909090909092e-06, 'num_tokens': 1485100.0, 'completions/mean_length': 123.25, 'completions/min_length': 63.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 67.71428680419922, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.8206631541252136, 'rewards/meter/std': 0.31621938943862915, '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/lexical_plausibility/mean': 0.9593336582183838, 'rewards/lexical_plausibility/std': 0.1150217279791832, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.7912750840187073, 'rewards/repeat_penalty/std': 0.15527135133743286, 'rewards/near_duplicate_penalty/mean': 0.8862694501876831, 'rewards/near_duplicate_penalty/std': 0.10902401059865952, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9205746054649353, 'rewards/distinct_2/std': 0.10719934105873108, 'rewards/total_composite/mean': 0.19626164436340332, 'rewards/total_composite/std': 0.16432058811187744, 'reward': 0.19626164436340332, 'reward_std': 0.16432060301303864, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1414690613746643, 'sampling/sampling_logp_difference/max': 1.495701789855957, 'sampling/importance_sampling_ratio/min': 0.2240912914276123, 'sampling/importance_sampling_ratio/mean': 1.0162965059280396, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9366847649216652, 'clip_ratio/low_mean': 0.06019270699471235, 'clip_ratio/low_min': 0.06019270699471235, 'clip_ratio/high_mean': 0.054251277819275856, 'clip_ratio/high_max': 0.054251277819275856, 'clip_ratio/region_mean': 0.11444398481398821, 'reward_total_mean': 0.19626164436340332, 'reward_meter_mean': 0.8206631541252136, 'reward_meter_std': 0.31621938943862915, '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_lexical_plausibility_mean': 0.9593336582183838, 'reward_lexical_plausibility_std': 0.1150217279791832, 'reward_repeat_penalty_mean': 0.7912750840187073, 'reward_repeat_penalty_std': 0.15527135133743286, 'reward_near_duplicate_penalty_mean': 0.8862694501876831, 'reward_near_duplicate_penalty_std': 0.10902401059865952, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9205746054649353, 'reward_distinct_2_std': 0.10719934105873108, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.19626164436340332, 'reward_total_composite_std': 0.16432058811187744, 'epoch': 0.03}

 22%|██▏       | 730/3300 [2:53:42<8:36:51, 12.07s/it]INFO 04-12 13:39:54 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 731/3300 [2:53:54<8:36:26, 12.06s/it]2026-04-12 13:40:06,089 | INFO | train_grpo_train | metrics_logged mode=train step=731

                                                      
{'loss': 0.0416, 'grad_norm': 7.710301876068115, 'learning_rate': 7.787878787878789e-06, 'num_tokens': 1488440.0, 'completions/mean_length': 221.5, 'completions/min_length': 175.0, 'completions/max_length': 271.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 221.5, 'completions/min_terminated_length': 175.0, 'completions/max_terminated_length': 271.0, 'rewards/meter/mean': 0.10196677595376968, 'rewards/meter/std': 0.04718315228819847, 'rewards/count_adherence/mean': 0.9722222089767456, 'rewards/count_adherence/std': 0.05143444985151291, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.5570678114891052, 'rewards/repeat_penalty/std': 0.3314151465892792, 'rewards/near_duplicate_penalty/mean': 0.9116718173027039, 'rewards/near_duplicate_penalty/std': 0.06395211815834045, 'rewards/opening_diversity/mean': 0.9791666269302368, 'rewards/opening_diversity/std': 0.0589255727827549, 'rewards/distinct_2/mean': 0.6194050312042236, 'rewards/distinct_2/std': 0.32400867342948914, 'rewards/total_composite/mean': 0.028781307861208916, 'rewards/total_composite/std': 0.017596958205103874, 'reward': 0.028781307861208916, 'reward_std': 0.017596958205103874, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13785415887832642, 'sampling/sampling_logp_difference/max': 3.0727105140686035, 'sampling/importance_sampling_ratio/min': 0.04629550129175186, 'sampling/importance_sampling_ratio/mean': 1.001731514930725, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5883641839027405, 'clip_ratio/low_mean': 0.053245002403855324, 'clip_ratio/low_min': 0.053245002403855324, 'clip_ratio/high_mean': 0.07681099325418472, 'clip_ratio/high_max': 0.07681099325418472, 'clip_ratio/region_mean': 0.13005599565804005, 'reward_total_mean': 0.028781307861208916, 'reward_meter_mean': 0.10196677595376968, 'reward_meter_std': 0.04718315228819847, 'reward_count_adherence_mean': 0.9722222089767456, 'reward_count_adherence_std': 0.05143444985151291, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.5570678114891052, 'reward_repeat_penalty_std': 0.3314151465892792, 'reward_near_duplicate_penalty_mean': 0.9116718173027039, 'reward_near_duplicate_penalty_std': 0.06395211815834045, 'reward_opening_diversity_mean': 0.9791666269302368, 'reward_opening_diversity_std': 0.0589255727827549, 'reward_distinct_2_mean': 0.6194050312042236, 'reward_distinct_2_std': 0.32400867342948914, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.028781307861208916, 'reward_total_composite_std': 0.017596958205103874, 'epoch': 0.03}

 22%|██▏       | 731/3300 [2:53:54<8:36:26, 12.06s/it]INFO 04-12 13:40:06 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 732/3300 [2:54:08<9:03:07, 12.69s/it]2026-04-12 13:40:20,242 | INFO | train_grpo_train | metrics_logged mode=train step=732

                                                      
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 7.784848484848484e-06, 'num_tokens': 1490208.0, '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': 0.9092035293579102, 'rewards/meter/std': 0.12226000428199768, 'rewards/count_adherence/mean': 0.9416667222976685, 'rewards/count_adherence/std': 0.055634863674640656, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.16250000894069672, 'rewards/judge_quality/std': 0.0353553369641304, 'rewards/repeat_penalty/mean': 0.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 0.4080020487308502, 'rewards/near_duplicate_penalty/std': 0.14329348504543304, 'rewards/opening_diversity/mean': 0.40406736731529236, 'rewards/opening_diversity/std': 0.3218400180339813, 'rewards/distinct_2/mean': 0.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.14159801602363586, 'rewards/total_composite/std': 0.0483369454741478, 'reward': 0.14159801602363586, 'reward_std': 0.0483369380235672, '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': 0.14159801602363586, 'reward_meter_mean': 0.9092035293579102, 'reward_meter_std': 0.12226000428199768, 'reward_count_adherence_mean': 0.9416667222976685, 'reward_count_adherence_std': 0.055634863674640656, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.0, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 0.4080020487308502, 'reward_near_duplicate_penalty_std': 0.14329348504543304, 'reward_opening_diversity_mean': 0.40406736731529236, 'reward_opening_diversity_std': 0.3218400180339813, 'reward_distinct_2_mean': 0.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.16250000894069672, 'reward_judge_quality_std': 0.0353553369641304, 'reward_total_composite_mean': 0.14159801602363586, 'reward_total_composite_std': 0.0483369454741478, 'epoch': 0.03}

 22%|██▏       | 732/3300 [2:54:08<9:03:07, 12.69s/it]INFO 04-12 13:40:20 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 733/3300 [2:54:21<9:13:02, 12.93s/it]2026-04-12 13:40:33,718 | INFO | train_grpo_train | metrics_logged mode=train step=733

                                                      
{'loss': -0.0322, 'grad_norm': 4.360764026641846, 'learning_rate': 7.781818181818183e-06, 'num_tokens': 1491859.0, 'completions/mean_length': 114.375, 'completions/min_length': 50.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 57.57143020629883, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.34302258491516113, 'rewards/meter/std': 0.38519468903541565, '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/lexical_plausibility/mean': 0.9680602550506592, 'rewards/lexical_plausibility/std': 0.0903393030166626, 'rewards/judge_quality/mean': 0.39625000953674316, 'rewards/judge_quality/std': 0.24720942974090576, 'rewards/repeat_penalty/mean': 0.9549835324287415, 'rewards/repeat_penalty/std': 0.024268770590424538, 'rewards/near_duplicate_penalty/mean': 0.9633446931838989, 'rewards/near_duplicate_penalty/std': 0.02702270820736885, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9916387796401978, 'rewards/distinct_2/std': 0.0236490610986948, 'rewards/total_composite/mean': 0.1877245455980301, 'rewards/total_composite/std': 0.28656256198883057, 'reward': 0.1877245455980301, 'reward_std': 0.2865625321865082, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14744533598423004, 'sampling/sampling_logp_difference/max': 1.5596216917037964, 'sampling/importance_sampling_ratio/min': 0.21021558344364166, 'sampling/importance_sampling_ratio/mean': 1.0043463706970215, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7598854750394821, 'clip_ratio/low_mean': 0.07134541124105453, 'clip_ratio/low_min': 0.07134541124105453, 'clip_ratio/high_mean': 0.04065604880452156, 'clip_ratio/high_max': 0.04065604880452156, 'clip_ratio/region_mean': 0.1120014600455761, 'reward_total_mean': 0.1877245455980301, 'reward_meter_mean': 0.34302258491516113, 'reward_meter_std': 0.38519468903541565, '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_lexical_plausibility_mean': 0.9680602550506592, 'reward_lexical_plausibility_std': 0.0903393030166626, 'reward_repeat_penalty_mean': 0.9549835324287415, 'reward_repeat_penalty_std': 0.024268770590424538, 'reward_near_duplicate_penalty_mean': 0.9633446931838989, 'reward_near_duplicate_penalty_std': 0.02702270820736885, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9916387796401978, 'reward_distinct_2_std': 0.0236490610986948, 'reward_judge_quality_mean': 0.39625000953674316, 'reward_judge_quality_std': 0.24720942974090576, 'reward_total_composite_mean': 0.1877245455980301, 'reward_total_composite_std': 0.28656256198883057, 'epoch': 0.03}

 22%|██▏       | 733/3300 [2:54:21<9:13:02, 12.93s/it]INFO 04-12 13:40:33 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 734/3300 [2:54:32<8:47:24, 12.33s/it]2026-04-12 13:40:44,664 | INFO | train_grpo_train | metrics_logged mode=train step=734

                                                      
{'loss': 0.0548, 'grad_norm': 8.31116771697998, 'learning_rate': 7.778787878787879e-06, 'num_tokens': 1494187.0, 'completions/mean_length': 111.0, 'completions/min_length': 105.0, 'completions/max_length': 118.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 111.0, 'completions/min_terminated_length': 105.0, 'completions/max_terminated_length': 118.0, 'rewards/meter/mean': 0.46877217292785645, 'rewards/meter/std': 0.310110479593277, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3375000059604645, 'rewards/judge_quality/std': 0.08345229178667068, 'rewards/repeat_penalty/mean': 0.9315893650054932, 'rewards/repeat_penalty/std': 0.04466979578137398, 'rewards/near_duplicate_penalty/mean': 0.96294766664505, 'rewards/near_duplicate_penalty/std': 0.014908126555383205, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9673303961753845, 'rewards/distinct_2/std': 0.04152150824666023, 'rewards/total_composite/mean': 0.14507833123207092, 'rewards/total_composite/std': 0.09024951606988907, 'reward': 0.14507833123207092, 'reward_std': 0.09024950861930847, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15470179915428162, 'sampling/sampling_logp_difference/max': 1.543485164642334, 'sampling/importance_sampling_ratio/min': 0.21363525092601776, 'sampling/importance_sampling_ratio/mean': 1.0147583484649658, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.055388405919075, 'clip_ratio/low_mean': 0.11407650262117386, 'clip_ratio/low_min': 0.11407650262117386, 'clip_ratio/high_mean': 0.05670710653066635, 'clip_ratio/high_max': 0.05670710653066635, 'clip_ratio/region_mean': 0.1707836091518402, 'reward_total_mean': 0.14507833123207092, 'reward_meter_mean': 0.46877217292785645, 'reward_meter_std': 0.310110479593277, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9315893650054932, 'reward_repeat_penalty_std': 0.04466979578137398, 'reward_near_duplicate_penalty_mean': 0.96294766664505, 'reward_near_duplicate_penalty_std': 0.014908126555383205, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9673303961753845, 'reward_distinct_2_std': 0.04152150824666023, 'reward_judge_quality_mean': 0.3375000059604645, 'reward_judge_quality_std': 0.08345229178667068, 'reward_total_composite_mean': 0.14507833123207092, 'reward_total_composite_std': 0.09024951606988907, 'epoch': 0.03}

 22%|██▏       | 734/3300 [2:54:32<8:47:24, 12.33s/it]INFO 04-12 13:40:44 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 735/3300 [2:54:40<7:54:31, 11.10s/it]2026-04-12 13:40:52,886 | INFO | train_grpo_train | metrics_logged mode=train step=735

                                                      
{'loss': 0.0235, 'grad_norm': 12.950492858886719, 'learning_rate': 7.775757575757576e-06, 'num_tokens': 1495778.0, 'completions/mean_length': 55.875, 'completions/min_length': 49.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.875, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.4620858430862427, 'rewards/meter/std': 0.34902605414390564, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4962500035762787, 'rewards/judge_quality/std': 0.17476005852222443, 'rewards/repeat_penalty/mean': 0.9603580236434937, 'rewards/repeat_penalty/std': 0.038065437227487564, 'rewards/near_duplicate_penalty/mean': 0.9695494174957275, 'rewards/near_duplicate_penalty/std': 0.0220591202378273, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.990384578704834, 'rewards/distinct_2/std': 0.027196412906050682, 'rewards/total_composite/mean': 0.20319171249866486, 'rewards/total_composite/std': 0.1373760998249054, 'reward': 0.20319171249866486, 'reward_std': 0.1373760998249054, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11961657553911209, 'sampling/sampling_logp_difference/max': 2.7539987564086914, 'sampling/importance_sampling_ratio/min': 0.063672736287117, 'sampling/importance_sampling_ratio/mean': 0.988656222820282, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5491676367819309, 'clip_ratio/low_mean': 0.06995941046625376, 'clip_ratio/low_min': 0.06995941046625376, 'clip_ratio/high_mean': 0.06452045869082212, 'clip_ratio/high_max': 0.06452045869082212, 'clip_ratio/region_mean': 0.13447986915707588, 'reward_total_mean': 0.20319171249866486, 'reward_meter_mean': 0.4620858430862427, 'reward_meter_std': 0.34902605414390564, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9603580236434937, 'reward_repeat_penalty_std': 0.038065437227487564, 'reward_near_duplicate_penalty_mean': 0.9695494174957275, 'reward_near_duplicate_penalty_std': 0.0220591202378273, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.990384578704834, 'reward_distinct_2_std': 0.027196412906050682, 'reward_judge_quality_mean': 0.4962500035762787, 'reward_judge_quality_std': 0.17476005852222443, 'reward_total_composite_mean': 0.20319171249866486, 'reward_total_composite_std': 0.1373760998249054, 'epoch': 0.03}

 22%|██▏       | 735/3300 [2:54:40<7:54:31, 11.10s/it]INFO 04-12 13:40:53 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 736/3300 [2:54:48<7:15:44, 10.20s/it]2026-04-12 13:41:00,980 | INFO | train_grpo_train | metrics_logged mode=train step=736

                                                      
{'loss': 0.0482, 'grad_norm': 10.444803237915039, 'learning_rate': 7.772727272727273e-06, 'num_tokens': 1497539.0, 'completions/mean_length': 64.125, 'completions/min_length': 54.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.125, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.7315667271614075, 'rewards/meter/std': 0.4078596532344818, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.12817399203777313, 'rewards/repeat_penalty/mean': 0.8645095825195312, 'rewards/repeat_penalty/std': 0.11634694784879684, 'rewards/near_duplicate_penalty/mean': 0.9057682752609253, 'rewards/near_duplicate_penalty/std': 0.03879751265048981, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9816433191299438, 'rewards/distinct_2/std': 0.03404124453663826, 'rewards/total_composite/mean': 0.25247079133987427, 'rewards/total_composite/std': 0.17357505857944489, 'reward': 0.25247079133987427, 'reward_std': 0.17357505857944489, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11564142256975174, 'sampling/sampling_logp_difference/max': 1.4629422426223755, 'sampling/importance_sampling_ratio/min': 0.23155400156974792, 'sampling/importance_sampling_ratio/mean': 1.0199613571166992, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8684811964631081, 'clip_ratio/low_mean': 0.05177712859585881, 'clip_ratio/low_min': 0.05177712859585881, 'clip_ratio/high_mean': 0.06499020010232925, 'clip_ratio/high_max': 0.06499020010232925, 'clip_ratio/region_mean': 0.11676732869818807, 'reward_total_mean': 0.25247079133987427, 'reward_meter_mean': 0.7315667271614075, 'reward_meter_std': 0.4078596532344818, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8645095825195312, 'reward_repeat_penalty_std': 0.11634694784879684, 'reward_near_duplicate_penalty_mean': 0.9057682752609253, 'reward_near_duplicate_penalty_std': 0.03879751265048981, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9816433191299438, 'reward_distinct_2_std': 0.03404124453663826, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.12817399203777313, 'reward_total_composite_mean': 0.25247079133987427, 'reward_total_composite_std': 0.17357505857944489, 'epoch': 0.03}

 22%|██▏       | 736/3300 [2:54:48<7:15:44, 10.20s/it]INFO 04-12 13:41:01 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 737/3300 [2:55:01<7:52:27, 11.06s/it]2026-04-12 13:41:14,052 | INFO | train_grpo_train | metrics_logged mode=train step=737

                                                      
{'loss': 0.0169, 'grad_norm': 3.542360305786133, 'learning_rate': 7.76969696969697e-06, 'num_tokens': 1499875.0, 'completions/mean_length': 205.0, 'completions/min_length': 124.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 161.1428680419922, 'completions/min_terminated_length': 124.0, 'completions/max_terminated_length': 180.0, 'rewards/meter/mean': 0.7619812488555908, 'rewards/meter/std': 0.34506988525390625, '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/lexical_plausibility/mean': 0.9356280565261841, 'rewards/lexical_plausibility/std': 0.18207144737243652, 'rewards/judge_quality/mean': 0.17500001192092896, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/repeat_penalty/mean': 0.3434876799583435, 'rewards/repeat_penalty/std': 0.33749091625213623, 'rewards/near_duplicate_penalty/mean': 0.7329788208007812, 'rewards/near_duplicate_penalty/std': 0.14577917754650116, 'rewards/opening_diversity/mean': 0.856249988079071, 'rewards/opening_diversity/std': 0.09425459057092667, 'rewards/distinct_2/mean': 0.4798237085342407, 'rewards/distinct_2/std': 0.38824746012687683, 'rewards/total_composite/mean': 0.11812707781791687, 'rewards/total_composite/std': 0.09984802454710007, 'reward': 0.11812707781791687, 'reward_std': 0.09984802454710007, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0892452746629715, 'sampling/sampling_logp_difference/max': 1.3417530059814453, 'sampling/importance_sampling_ratio/min': 0.26138705015182495, 'sampling/importance_sampling_ratio/mean': 1.0022655725479126, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5627057887613773, 'clip_ratio/low_mean': 0.05400221236050129, 'clip_ratio/low_min': 0.05400221236050129, 'clip_ratio/high_mean': 0.016129031777381897, 'clip_ratio/high_max': 0.016129031777381897, 'clip_ratio/region_mean': 0.07013124413788319, 'reward_total_mean': 0.11812707781791687, 'reward_meter_mean': 0.7619812488555908, 'reward_meter_std': 0.34506988525390625, '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_lexical_plausibility_mean': 0.9356280565261841, 'reward_lexical_plausibility_std': 0.18207144737243652, 'reward_repeat_penalty_mean': 0.3434876799583435, 'reward_repeat_penalty_std': 0.33749091625213623, 'reward_near_duplicate_penalty_mean': 0.7329788208007812, 'reward_near_duplicate_penalty_std': 0.14577917754650116, 'reward_opening_diversity_mean': 0.856249988079071, 'reward_opening_diversity_std': 0.09425459057092667, 'reward_distinct_2_mean': 0.4798237085342407, 'reward_distinct_2_std': 0.38824746012687683, 'reward_judge_quality_mean': 0.17500001192092896, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.11812707781791687, 'reward_total_composite_std': 0.09984802454710007, 'epoch': 0.03}

 22%|██▏       | 737/3300 [2:55:02<7:52:27, 11.06s/it]INFO 04-12 13:41:14 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 738/3300 [2:55:15<8:27:21, 11.88s/it]2026-04-12 13:41:27,851 | INFO | train_grpo_train | metrics_logged mode=train step=738

                                                      
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 7.766666666666666e-06, 'num_tokens': 1501499.0, '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': 0.8967434763908386, 'rewards/meter/std': 0.18547113239765167, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.09155285358428955, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.996417760848999, 'rewards/lexical_plausibility/std': 0.010131997056305408, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.003861095989122987, 'rewards/repeat_penalty/std': 0.010920829139649868, 'rewards/near_duplicate_penalty/mean': 0.5388668775558472, 'rewards/near_duplicate_penalty/std': 0.19304721057415009, 'rewards/opening_diversity/mean': 0.7080065011978149, 'rewards/opening_diversity/std': 0.2737760841846466, 'rewards/distinct_2/mean': 0.006916329264640808, 'rewards/distinct_2/std': 0.019562333822250366, 'rewards/total_composite/mean': 0.11902313679456711, 'rewards/total_composite/std': 0.031959667801856995, 'reward': 0.11902313679456711, 'reward_std': 0.031959664076566696, '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': 0.11902313679456711, 'reward_meter_mean': 0.8967434763908386, 'reward_meter_std': 0.18547113239765167, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.09155285358428955, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.996417760848999, 'reward_lexical_plausibility_std': 0.010131997056305408, 'reward_repeat_penalty_mean': 0.003861095989122987, 'reward_repeat_penalty_std': 0.010920829139649868, 'reward_near_duplicate_penalty_mean': 0.5388668775558472, 'reward_near_duplicate_penalty_std': 0.19304721057415009, 'reward_opening_diversity_mean': 0.7080065011978149, 'reward_opening_diversity_std': 0.2737760841846466, 'reward_distinct_2_mean': 0.006916329264640808, 'reward_distinct_2_std': 0.019562333822250366, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.11902313679456711, 'reward_total_composite_std': 0.031959667801856995, 'epoch': 0.03}

 22%|██▏       | 738/3300 [2:55:15<8:27:21, 11.88s/it]INFO 04-12 13:41:28 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 739/3300 [2:55:27<8:19:15, 11.70s/it]2026-04-12 13:41:39,116 | INFO | train_grpo_train | metrics_logged mode=train step=739

                                                      
{'loss': -0.0003, 'grad_norm': 4.201359748840332, 'learning_rate': 7.763636363636364e-06, 'num_tokens': 1505185.0, 'completions/mean_length': 280.75, 'completions/min_length': 250.0, 'completions/max_length': 320.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 280.75, 'completions/min_terminated_length': 250.0, 'completions/max_terminated_length': 320.0, 'rewards/meter/mean': 0.7154779434204102, 'rewards/meter/std': 0.4361130893230438, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.22500000894069672, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.040792446583509445, 'rewards/repeat_penalty/std': 0.056984249502420425, 'rewards/near_duplicate_penalty/mean': 0.471712589263916, 'rewards/near_duplicate_penalty/std': 0.19876046478748322, 'rewards/opening_diversity/mean': 0.7401316165924072, 'rewards/opening_diversity/std': 0.2737102210521698, 'rewards/distinct_2/mean': 0.10464771836996078, 'rewards/distinct_2/std': 0.14465248584747314, 'rewards/total_composite/mean': 0.13055403530597687, 'rewards/total_composite/std': 0.10409002006053925, 'reward': 0.13055403530597687, 'reward_std': 0.10409002006053925, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05282207950949669, 'sampling/sampling_logp_difference/max': 1.8779902458190918, 'sampling/importance_sampling_ratio/min': 0.15289708971977234, 'sampling/importance_sampling_ratio/mean': 1.0042352676391602, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.29088243283331394, 'clip_ratio/low_mean': 0.022267282009124756, 'clip_ratio/low_min': 0.022267282009124756, 'clip_ratio/high_mean': 0.02629257389344275, 'clip_ratio/high_max': 0.02629257389344275, 'clip_ratio/region_mean': 0.048559855902567506, 'reward_total_mean': 0.13055403530597687, 'reward_meter_mean': 0.7154779434204102, 'reward_meter_std': 0.4361130893230438, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.040792446583509445, 'reward_repeat_penalty_std': 0.056984249502420425, 'reward_near_duplicate_penalty_mean': 0.471712589263916, 'reward_near_duplicate_penalty_std': 0.19876046478748322, 'reward_opening_diversity_mean': 0.7401316165924072, 'reward_opening_diversity_std': 0.2737102210521698, 'reward_distinct_2_mean': 0.10464771836996078, 'reward_distinct_2_std': 0.14465248584747314, 'reward_judge_quality_mean': 0.22500000894069672, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.13055403530597687, 'reward_total_composite_std': 0.10409002006053925, 'epoch': 0.03}

 22%|██▏       | 739/3300 [2:55:27<8:19:15, 11.70s/it]INFO 04-12 13:41:39 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 740/3300 [2:55:41<8:50:38, 12.44s/it]2026-04-12 13:41:53,285 | INFO | train_grpo_train | metrics_logged mode=train step=740

                                                      
{'loss': -0.1977, 'grad_norm': 2.4809043407440186, 'learning_rate': 7.76060606060606e-06, 'num_tokens': 1507258.0, 'completions/mean_length': 272.125, 'completions/min_length': 110.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 128.1999969482422, 'completions/min_terminated_length': 110.0, 'completions/max_terminated_length': 155.0, 'rewards/meter/mean': 0.8252730369567871, 'rewards/meter/std': 0.2637774348258972, '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/lexical_plausibility/mean': 0.8784559965133667, 'rewards/lexical_plausibility/std': 0.17994637787342072, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.16903084516525269, 'rewards/repeat_penalty/mean': 0.9217922687530518, 'rewards/repeat_penalty/std': 0.07210688292980194, 'rewards/near_duplicate_penalty/mean': 0.9669778943061829, 'rewards/near_duplicate_penalty/std': 0.02374645136296749, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9525356888771057, 'rewards/distinct_2/std': 0.05886922776699066, 'rewards/total_composite/mean': 0.21614933013916016, 'rewards/total_composite/std': 0.1745433509349823, 'reward': 0.21614933013916016, 'reward_std': 0.1745433360338211, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12845279276371002, 'sampling/sampling_logp_difference/max': 1.1226415634155273, 'sampling/importance_sampling_ratio/min': 0.32541903853416443, 'sampling/importance_sampling_ratio/mean': 1.0235168933868408, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7834031879901886, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.09019136149436235, 'clip_ratio/high_max': 0.09019136149436235, 'clip_ratio/region_mean': 0.09019136149436235, 'reward_total_mean': 0.21614933013916016, 'reward_meter_mean': 0.8252730369567871, 'reward_meter_std': 0.2637774348258972, '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_lexical_plausibility_mean': 0.8784559965133667, 'reward_lexical_plausibility_std': 0.17994637787342072, 'reward_repeat_penalty_mean': 0.9217922687530518, 'reward_repeat_penalty_std': 0.07210688292980194, 'reward_near_duplicate_penalty_mean': 0.9669778943061829, 'reward_near_duplicate_penalty_std': 0.02374645136296749, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9525356888771057, 'reward_distinct_2_std': 0.05886922776699066, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.16903084516525269, 'reward_total_composite_mean': 0.21614933013916016, 'reward_total_composite_std': 0.1745433509349823, 'epoch': 0.03}

 22%|██▏       | 740/3300 [2:55:41<8:50:38, 12.44s/it]INFO 04-12 13:41:53 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 741/3300 [2:55:54<8:58:36, 12.63s/it]2026-04-12 13:42:06,352 | INFO | train_grpo_train | metrics_logged mode=train step=741

                                                      
{'loss': -0.0781, 'grad_norm': 3.3963463306427, 'learning_rate': 7.757575757575758e-06, 'num_tokens': 1509167.0, 'completions/mean_length': 195.625, 'completions/min_length': 82.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 90.16667175292969, 'completions/min_terminated_length': 82.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.4642590284347534, 'rewards/meter/std': 0.3871473968029022, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.8788434863090515, 'rewards/lexical_plausibility/std': 0.1749013513326645, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.155264750123024, 'rewards/repeat_penalty/mean': 0.9139944314956665, 'rewards/repeat_penalty/std': 0.08927091956138611, 'rewards/near_duplicate_penalty/mean': 0.962729275226593, 'rewards/near_duplicate_penalty/std': 0.03199862316250801, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9479074478149414, 'rewards/distinct_2/std': 0.06861712038516998, 'rewards/total_composite/mean': 0.11039489507675171, 'rewards/total_composite/std': 0.12055056542158127, 'reward': 0.11039489507675171, 'reward_std': 0.12055056542158127, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11214577406644821, 'sampling/sampling_logp_difference/max': 2.559469699859619, 'sampling/importance_sampling_ratio/min': 0.07734574377536774, 'sampling/importance_sampling_ratio/mean': 1.0267283916473389, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.49914003163576126, 'clip_ratio/low_mean': 0.03309566364623606, 'clip_ratio/low_min': 0.03309566364623606, 'clip_ratio/high_mean': 0.042008678428828716, 'clip_ratio/high_max': 0.042008678428828716, 'clip_ratio/region_mean': 0.07510434207506478, 'reward_total_mean': 0.11039489507675171, 'reward_meter_mean': 0.4642590284347534, 'reward_meter_std': 0.3871473968029022, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.8788434863090515, 'reward_lexical_plausibility_std': 0.1749013513326645, 'reward_repeat_penalty_mean': 0.9139944314956665, 'reward_repeat_penalty_std': 0.08927091956138611, 'reward_near_duplicate_penalty_mean': 0.962729275226593, 'reward_near_duplicate_penalty_std': 0.03199862316250801, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9479074478149414, 'reward_distinct_2_std': 0.06861712038516998, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.155264750123024, 'reward_total_composite_mean': 0.11039489507675171, 'reward_total_composite_std': 0.12055056542158127, 'epoch': 0.03}

 22%|██▏       | 741/3300 [2:55:54<8:58:36, 12.63s/it]INFO 04-12 13:42:06 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 742/3300 [2:56:07<9:08:05, 12.86s/it]2026-04-12 13:42:19,742 | INFO | train_grpo_train | metrics_logged mode=train step=742

                                                      
{'loss': -0.1242, 'grad_norm': 4.418141841888428, 'learning_rate': 7.754545454545455e-06, 'num_tokens': 1511395.0, 'completions/mean_length': 172.5, 'completions/min_length': 119.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 124.00000762939453, 'completions/min_terminated_length': 119.0, 'completions/max_terminated_length': 135.0, 'rewards/meter/mean': 0.783625602722168, 'rewards/meter/std': 0.3340771794319153, '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/lexical_plausibility/mean': 0.9600352048873901, 'rewards/lexical_plausibility/std': 0.11303751915693283, 'rewards/judge_quality/mean': 0.2562500238418579, 'rewards/judge_quality/std': 0.13741880655288696, 'rewards/repeat_penalty/mean': 0.6281116008758545, 'rewards/repeat_penalty/std': 0.3001388907432556, 'rewards/near_duplicate_penalty/mean': 0.8636089563369751, 'rewards/near_duplicate_penalty/std': 0.10903690755367279, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1530931144952774, 'rewards/distinct_2/mean': 0.7256183624267578, 'rewards/distinct_2/std': 0.26374635100364685, 'rewards/total_composite/mean': 0.21634042263031006, 'rewards/total_composite/std': 0.13507242500782013, 'reward': 0.21634042263031006, 'reward_std': 0.13507241010665894, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1075095534324646, 'sampling/sampling_logp_difference/max': 1.510689616203308, 'sampling/importance_sampling_ratio/min': 0.22075767815113068, 'sampling/importance_sampling_ratio/mean': 1.0034854412078857, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6330233700573444, 'clip_ratio/low_mean': 0.0325851459056139, 'clip_ratio/low_min': 0.0325851459056139, 'clip_ratio/high_mean': 0.06569452211260796, 'clip_ratio/high_max': 0.06569452211260796, 'clip_ratio/region_mean': 0.09827966801822186, 'reward_total_mean': 0.21634042263031006, 'reward_meter_mean': 0.783625602722168, 'reward_meter_std': 0.3340771794319153, '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_lexical_plausibility_mean': 0.9600352048873901, 'reward_lexical_plausibility_std': 0.11303751915693283, 'reward_repeat_penalty_mean': 0.6281116008758545, 'reward_repeat_penalty_std': 0.3001388907432556, 'reward_near_duplicate_penalty_mean': 0.8636089563369751, 'reward_near_duplicate_penalty_std': 0.10903690755367279, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1530931144952774, 'reward_distinct_2_mean': 0.7256183624267578, 'reward_distinct_2_std': 0.26374635100364685, 'reward_judge_quality_mean': 0.2562500238418579, 'reward_judge_quality_std': 0.13741880655288696, 'reward_total_composite_mean': 0.21634042263031006, 'reward_total_composite_std': 0.13507242500782013, 'epoch': 0.03}

 22%|██▏       | 742/3300 [2:56:07<9:08:05, 12.86s/it]INFO 04-12 13:42:19 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 743/3300 [2:56:21<9:20:50, 13.16s/it]2026-04-12 13:42:33,610 | INFO | train_grpo_train | metrics_logged mode=train step=743

                                                      
{'loss': -0.0985, 'grad_norm': 1.9212191104888916, 'learning_rate': 7.751515151515153e-06, 'num_tokens': 1514982.0, 'completions/mean_length': 485.375, 'completions/min_length': 423.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 458.75, 'completions/min_terminated_length': 423.0, 'completions/max_terminated_length': 503.0, 'rewards/meter/mean': 0.5891445875167847, 'rewards/meter/std': 0.4586457908153534, 'rewards/count_adherence/mean': 0.759615421295166, 'rewards/count_adherence/std': 0.06419406831264496, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.012268980965018272, 'rewards/repeat_penalty/std': 0.022723762318491936, 'rewards/near_duplicate_penalty/mean': 0.5857747793197632, 'rewards/near_duplicate_penalty/std': 0.14130906760692596, 'rewards/opening_diversity/mean': 0.7407532334327698, 'rewards/opening_diversity/std': 0.27918604016304016, 'rewards/distinct_2/mean': 0.023685414344072342, 'rewards/distinct_2/std': 0.04385910555720329, 'rewards/total_composite/mean': 0.06668651849031448, 'rewards/total_composite/std': 0.05276230350136757, 'reward': 0.06668651849031448, 'reward_std': 0.05276230350136757, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.044047191739082336, 'sampling/sampling_logp_difference/max': 2.1677281856536865, 'sampling/importance_sampling_ratio/min': 0.11443730443716049, 'sampling/importance_sampling_ratio/mean': 1.0064189434051514, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.11544675100594759, 'clip_ratio/low_mean': 0.007336343172937632, 'clip_ratio/low_min': 0.007336343172937632, 'clip_ratio/high_mean': 0.011733844876289368, 'clip_ratio/high_max': 0.011733844876289368, 'clip_ratio/region_mean': 0.019070188049227, 'reward_total_mean': 0.06668651849031448, 'reward_meter_mean': 0.5891445875167847, 'reward_meter_std': 0.4586457908153534, 'reward_count_adherence_mean': 0.759615421295166, 'reward_count_adherence_std': 0.06419406831264496, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.012268980965018272, 'reward_repeat_penalty_std': 0.022723762318491936, 'reward_near_duplicate_penalty_mean': 0.5857747793197632, 'reward_near_duplicate_penalty_std': 0.14130906760692596, 'reward_opening_diversity_mean': 0.7407532334327698, 'reward_opening_diversity_std': 0.27918604016304016, 'reward_distinct_2_mean': 0.023685414344072342, 'reward_distinct_2_std': 0.04385910555720329, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.06668651849031448, 'reward_total_composite_std': 0.05276230350136757, 'epoch': 0.03}

 23%|██▎       | 743/3300 [2:56:21<9:20:50, 13.16s/it]INFO 04-12 13:42:33 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 744/3300 [2:56:37<9:52:31, 13.91s/it]2026-04-12 13:42:49,268 | INFO | train_grpo_train | metrics_logged mode=train step=744

                                                      
{'loss': -0.0357, 'grad_norm': 5.227776050567627, 'learning_rate': 7.74848484848485e-06, 'num_tokens': 1516631.0, 'completions/mean_length': 111.125, 'completions/min_length': 45.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 53.85714340209961, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.5472466349601746, 'rewards/meter/std': 0.4355945885181427, '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/lexical_plausibility/mean': 0.9639897346496582, 'rewards/lexical_plausibility/std': 0.10185245424509048, 'rewards/judge_quality/mean': 0.4337500035762787, 'rewards/judge_quality/std': 0.30677762627601624, 'rewards/repeat_penalty/mean': 0.9398870468139648, 'rewards/repeat_penalty/std': 0.032975081354379654, 'rewards/near_duplicate_penalty/mean': 0.9404217004776001, 'rewards/near_duplicate_penalty/std': 0.03266766294836998, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9994224905967712, 'rewards/distinct_2/std': 0.001633422332815826, 'rewards/total_composite/mean': 0.23873844742774963, 'rewards/total_composite/std': 0.278200626373291, 'reward': 0.23873844742774963, 'reward_std': 0.278200626373291, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1565236896276474, 'sampling/sampling_logp_difference/max': 1.5941534042358398, 'sampling/importance_sampling_ratio/min': 0.20308038592338562, 'sampling/importance_sampling_ratio/mean': 1.0218241214752197, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9597418680787086, 'clip_ratio/low_mean': 0.07491765636950731, 'clip_ratio/low_min': 0.07491765636950731, 'clip_ratio/high_mean': 0.0514801568351686, 'clip_ratio/high_max': 0.0514801568351686, 'clip_ratio/region_mean': 0.1263978132046759, 'reward_total_mean': 0.23873844742774963, 'reward_meter_mean': 0.5472466349601746, 'reward_meter_std': 0.4355945885181427, '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_lexical_plausibility_mean': 0.9639897346496582, 'reward_lexical_plausibility_std': 0.10185245424509048, 'reward_repeat_penalty_mean': 0.9398870468139648, 'reward_repeat_penalty_std': 0.032975081354379654, 'reward_near_duplicate_penalty_mean': 0.9404217004776001, 'reward_near_duplicate_penalty_std': 0.03266766294836998, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9994224905967712, 'reward_distinct_2_std': 0.001633422332815826, 'reward_judge_quality_mean': 0.4337500035762787, 'reward_judge_quality_std': 0.30677762627601624, 'reward_total_composite_mean': 0.23873844742774963, 'reward_total_composite_std': 0.278200626373291, 'epoch': 0.03}

 23%|██▎       | 744/3300 [2:56:37<9:52:31, 13.91s/it]INFO 04-12 13:42:49 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 745/3300 [2:56:45<8:44:44, 12.32s/it]2026-04-12 13:42:57,889 | INFO | train_grpo_train | metrics_logged mode=train step=745

                                                      
{'loss': 0.0391, 'grad_norm': 11.293991088867188, 'learning_rate': 7.745454545454545e-06, 'num_tokens': 1518327.0, 'completions/mean_length': 55.0, 'completions/min_length': 52.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.0, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.9827555418014526, 'rewards/meter/std': 0.016169028356671333, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6000000238418579, 'rewards/judge_quality/std': 0.2317017912864685, 'rewards/repeat_penalty/mean': 0.9730241298675537, 'rewards/repeat_penalty/std': 0.022936878725886345, 'rewards/near_duplicate_penalty/mean': 0.9730241298675537, 'rewards/near_duplicate_penalty/std': 0.022936878725886345, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5882487297058105, 'rewards/total_composite/std': 0.22248080372810364, 'reward': 0.5882487297058105, 'reward_std': 0.22248077392578125, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12161202728748322, 'sampling/sampling_logp_difference/max': 1.2406206130981445, 'sampling/importance_sampling_ratio/min': 0.2892046868801117, 'sampling/importance_sampling_ratio/mean': 1.0001301765441895, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8677456304430962, 'clip_ratio/low_mean': 0.0786271607503295, 'clip_ratio/low_min': 0.0786271607503295, 'clip_ratio/high_mean': 0.04597964696586132, 'clip_ratio/high_max': 0.04597964696586132, 'clip_ratio/region_mean': 0.12460680771619081, 'reward_total_mean': 0.5882487297058105, 'reward_meter_mean': 0.9827555418014526, 'reward_meter_std': 0.016169028356671333, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9730241298675537, 'reward_repeat_penalty_std': 0.022936878725886345, 'reward_near_duplicate_penalty_mean': 0.9730241298675537, 'reward_near_duplicate_penalty_std': 0.022936878725886345, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6000000238418579, 'reward_judge_quality_std': 0.2317017912864685, 'reward_total_composite_mean': 0.5882487297058105, 'reward_total_composite_std': 0.22248080372810364, 'epoch': 0.03}

 23%|██▎       | 745/3300 [2:56:45<8:44:44, 12.32s/it]INFO 04-12 13:42:58 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 746/3300 [2:56:55<8:14:30, 11.62s/it]2026-04-12 13:43:07,864 | INFO | train_grpo_train | metrics_logged mode=train step=746

                                                      
{'loss': 0.0331, 'grad_norm': 6.199829578399658, 'learning_rate': 7.742424242424244e-06, 'num_tokens': 1521104.0, 'completions/mean_length': 157.125, 'completions/min_length': 144.0, 'completions/max_length': 176.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 157.125, 'completions/min_terminated_length': 144.0, 'completions/max_terminated_length': 176.0, 'rewards/meter/mean': 0.7413284778594971, 'rewards/meter/std': 0.38702788949012756, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.7569299936294556, 'rewards/repeat_penalty/std': 0.13841694593429565, 'rewards/near_duplicate_penalty/mean': 0.9138485789299011, 'rewards/near_duplicate_penalty/std': 0.04501912370324135, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.823939859867096, 'rewards/distinct_2/std': 0.11312751471996307, 'rewards/total_composite/mean': 0.2502559721469879, 'rewards/total_composite/std': 0.14552713930606842, 'reward': 0.2502559721469879, 'reward_std': 0.14552712440490723, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1081324964761734, 'sampling/sampling_logp_difference/max': 1.7347493171691895, 'sampling/importance_sampling_ratio/min': 0.1764444261789322, 'sampling/importance_sampling_ratio/mean': 1.0021672248840332, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5856231823563576, 'clip_ratio/low_mean': 0.03457912337034941, 'clip_ratio/low_min': 0.03457912337034941, 'clip_ratio/high_mean': 0.07685709651559591, 'clip_ratio/high_max': 0.07685709651559591, 'clip_ratio/region_mean': 0.11143621988594532, 'reward_total_mean': 0.2502559721469879, 'reward_meter_mean': 0.7413284778594971, 'reward_meter_std': 0.38702788949012756, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7569299936294556, 'reward_repeat_penalty_std': 0.13841694593429565, 'reward_near_duplicate_penalty_mean': 0.9138485789299011, 'reward_near_duplicate_penalty_std': 0.04501912370324135, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.823939859867096, 'reward_distinct_2_std': 0.11312751471996307, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.2502559721469879, 'reward_total_composite_std': 0.14552713930606842, 'epoch': 0.03}

 23%|██▎       | 746/3300 [2:56:55<8:14:30, 11.62s/it]INFO 04-12 13:43:08 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 747/3300 [2:57:08<8:31:09, 12.01s/it]2026-04-12 13:43:20,796 | INFO | train_grpo_train | metrics_logged mode=train step=747

                                                      
{'loss': -0.0377, 'grad_norm': 2.940369129180908, 'learning_rate': 7.73939393939394e-06, 'num_tokens': 1522478.0, 'completions/mean_length': 152.75, 'completions/min_length': 29.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 33.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.7009673118591309, 'rewards/meter/std': 0.4046489894390106, '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/lexical_plausibility/mean': 0.8916381597518921, 'rewards/lexical_plausibility/std': 0.16471542418003082, 'rewards/judge_quality/mean': 0.3487499952316284, 'rewards/judge_quality/std': 0.36883747577667236, 'rewards/repeat_penalty/mean': 0.7422963976860046, 'rewards/repeat_penalty/std': 0.1507778763771057, 'rewards/near_duplicate_penalty/mean': 0.9631261825561523, 'rewards/near_duplicate_penalty/std': 0.10429497063159943, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9786324501037598, 'rewards/distinct_2/std': 0.06043647602200508, 'rewards/total_composite/mean': 0.2553945779800415, 'rewards/total_composite/std': 0.2871827483177185, 'reward': 0.2553945779800415, 'reward_std': 0.2871827483177185, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1165352463722229, 'sampling/sampling_logp_difference/max': 0.9058531522750854, 'sampling/importance_sampling_ratio/min': 0.40419691801071167, 'sampling/importance_sampling_ratio/mean': 1.0442841053009033, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6533161997795105, 'clip_ratio/low_mean': 0.03984700050204992, 'clip_ratio/low_min': 0.03984700050204992, 'clip_ratio/high_mean': 0.05170829128473997, 'clip_ratio/high_max': 0.05170829128473997, 'clip_ratio/region_mean': 0.0915552917867899, 'reward_total_mean': 0.2553945779800415, 'reward_meter_mean': 0.7009673118591309, 'reward_meter_std': 0.4046489894390106, '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_lexical_plausibility_mean': 0.8916381597518921, 'reward_lexical_plausibility_std': 0.16471542418003082, 'reward_repeat_penalty_mean': 0.7422963976860046, 'reward_repeat_penalty_std': 0.1507778763771057, 'reward_near_duplicate_penalty_mean': 0.9631261825561523, 'reward_near_duplicate_penalty_std': 0.10429497063159943, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9786324501037598, 'reward_distinct_2_std': 0.06043647602200508, 'reward_judge_quality_mean': 0.3487499952316284, 'reward_judge_quality_std': 0.36883747577667236, 'reward_total_composite_mean': 0.2553945779800415, 'reward_total_composite_std': 0.2871827483177185, 'epoch': 0.03}

 23%|██▎       | 747/3300 [2:57:08<8:31:09, 12.01s/it]INFO 04-12 13:43:21 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 748/3300 [2:57:17<7:46:48, 10.98s/it]2026-04-12 13:43:29,352 | INFO | train_grpo_train | metrics_logged mode=train step=748

                                                      
{'loss': -0.0053, 'grad_norm': 10.388566017150879, 'learning_rate': 7.736363636363637e-06, 'num_tokens': 1524163.0, 'completions/mean_length': 57.625, 'completions/min_length': 49.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.625, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9111192226409912, 'rewards/meter/std': 0.19375477731227875, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4962499737739563, 'rewards/judge_quality/std': 0.17476005852222443, 'rewards/repeat_penalty/mean': 0.9604243040084839, 'rewards/repeat_penalty/std': 0.057449083775281906, 'rewards/near_duplicate_penalty/mean': 0.9687251448631287, 'rewards/near_duplicate_penalty/std': 0.039593104273080826, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9908424615859985, 'rewards/distinct_2/std': 0.025901345536112785, 'rewards/total_composite/mean': 0.45587459206581116, 'rewards/total_composite/std': 0.20374882221221924, 'reward': 0.45587459206581116, 'reward_std': 0.20374882221221924, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11318694800138474, 'sampling/sampling_logp_difference/max': 1.5116311311721802, 'sampling/importance_sampling_ratio/min': 0.22054994106292725, 'sampling/importance_sampling_ratio/mean': 1.0068858861923218, 'sampling/importance_sampling_ratio/max': 1.795382022857666, 'entropy': 0.6656927317380905, 'clip_ratio/low_mean': 0.077200872823596, 'clip_ratio/low_min': 0.077200872823596, 'clip_ratio/high_mean': 0.004385964944958687, 'clip_ratio/high_max': 0.004385964944958687, 'clip_ratio/region_mean': 0.08158683776855469, 'reward_total_mean': 0.45587459206581116, 'reward_meter_mean': 0.9111192226409912, 'reward_meter_std': 0.19375477731227875, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9604243040084839, 'reward_repeat_penalty_std': 0.057449083775281906, 'reward_near_duplicate_penalty_mean': 0.9687251448631287, 'reward_near_duplicate_penalty_std': 0.039593104273080826, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9908424615859985, 'reward_distinct_2_std': 0.025901345536112785, 'reward_judge_quality_mean': 0.4962499737739563, 'reward_judge_quality_std': 0.17476005852222443, 'reward_total_composite_mean': 0.45587459206581116, 'reward_total_composite_std': 0.20374882221221924, 'epoch': 0.03}

 23%|██▎       | 748/3300 [2:57:17<7:46:48, 10.98s/it]INFO 04-12 13:43:29 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 749/3300 [2:57:30<8:16:41, 11.68s/it]2026-04-12 13:43:42,682 | INFO | train_grpo_train | metrics_logged mode=train step=749

                                                      
{'loss': -0.0712, 'grad_norm': 5.339338302612305, 'learning_rate': 7.733333333333334e-06, 'num_tokens': 1526061.0, 'completions/mean_length': 137.25, 'completions/min_length': 77.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 83.71428680419922, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 91.0, 'rewards/meter/mean': 0.6527518630027771, 'rewards/meter/std': 0.39722979068756104, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.932406485080719, 'rewards/lexical_plausibility/std': 0.19118337333202362, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.1414213478565216, 'rewards/repeat_penalty/mean': 0.892428457736969, 'rewards/repeat_penalty/std': 0.08709447830915451, 'rewards/near_duplicate_penalty/mean': 0.9321856498718262, 'rewards/near_duplicate_penalty/std': 0.0488106794655323, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9555417895317078, 'rewards/distinct_2/std': 0.05076823756098747, 'rewards/total_composite/mean': 0.23975013196468353, 'rewards/total_composite/std': 0.1801433563232422, 'reward': 0.23975013196468353, 'reward_std': 0.1801433563232422, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1256818324327469, 'sampling/sampling_logp_difference/max': 1.4742236137390137, 'sampling/importance_sampling_ratio/min': 0.2289564162492752, 'sampling/importance_sampling_ratio/mean': 1.0139474868774414, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6879321783781052, 'clip_ratio/low_mean': 0.029454919509589672, 'clip_ratio/low_min': 0.029454919509589672, 'clip_ratio/high_mean': 0.0764033542945981, 'clip_ratio/high_max': 0.0764033542945981, 'clip_ratio/region_mean': 0.10585827380418777, 'reward_total_mean': 0.23975013196468353, 'reward_meter_mean': 0.6527518630027771, 'reward_meter_std': 0.39722979068756104, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.932406485080719, 'reward_lexical_plausibility_std': 0.19118337333202362, 'reward_repeat_penalty_mean': 0.892428457736969, 'reward_repeat_penalty_std': 0.08709447830915451, 'reward_near_duplicate_penalty_mean': 0.9321856498718262, 'reward_near_duplicate_penalty_std': 0.0488106794655323, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9555417895317078, 'reward_distinct_2_std': 0.05076823756098747, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.1414213478565216, 'reward_total_composite_mean': 0.23975013196468353, 'reward_total_composite_std': 0.1801433563232422, 'epoch': 0.03}

 23%|██▎       | 749/3300 [2:57:30<8:16:41, 11.68s/it]INFO 04-12 13:43:42 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 750/3300 [2:57:40<7:55:05, 11.18s/it]2026-04-12 13:43:52,685 | INFO | train_grpo_train | metrics_logged mode=train step=750

                                                      
{'loss': 0.0245, 'grad_norm': 11.843229293823242, 'learning_rate': 7.730303030303032e-06, 'num_tokens': 1527911.0, 'completions/mean_length': 62.25, 'completions/min_length': 58.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.25, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.6764695644378662, 'rewards/meter/std': 0.412853866815567, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9623015522956848, 'rewards/lexical_plausibility/std': 0.06398618221282959, 'rewards/judge_quality/mean': 0.42499998211860657, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.9751977324485779, 'rewards/repeat_penalty/std': 0.03601844236254692, 'rewards/near_duplicate_penalty/mean': 0.9851332306861877, 'rewards/near_duplicate_penalty/std': 0.020686792209744453, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9898785352706909, 'rewards/distinct_2/std': 0.028627805411815643, 'rewards/total_composite/mean': 0.29100680351257324, 'rewards/total_composite/std': 0.17981988191604614, 'reward': 0.29100680351257324, 'reward_std': 0.17981988191604614, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.169460266828537, 'sampling/sampling_logp_difference/max': 1.4786615371704102, 'sampling/importance_sampling_ratio/min': 0.2279425859451294, 'sampling/importance_sampling_ratio/mean': 1.0167688131332397, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0941259264945984, 'clip_ratio/low_mean': 0.056163594126701355, 'clip_ratio/low_min': 0.056163594126701355, 'clip_ratio/high_mean': 0.08973436057567596, 'clip_ratio/high_max': 0.08973436057567596, 'clip_ratio/region_mean': 0.14589795470237732, 'reward_total_mean': 0.29100680351257324, 'reward_meter_mean': 0.6764695644378662, 'reward_meter_std': 0.412853866815567, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9623015522956848, 'reward_lexical_plausibility_std': 0.06398618221282959, 'reward_repeat_penalty_mean': 0.9751977324485779, 'reward_repeat_penalty_std': 0.03601844236254692, 'reward_near_duplicate_penalty_mean': 0.9851332306861877, 'reward_near_duplicate_penalty_std': 0.020686792209744453, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9898785352706909, 'reward_distinct_2_std': 0.028627805411815643, 'reward_judge_quality_mean': 0.42499998211860657, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.29100680351257324, 'reward_total_composite_std': 0.17981988191604614, 'epoch': 0.03}

 23%|██▎       | 750/3300 [2:57:40<7:55:05, 11.18s/it]INFO 04-12 13:43:52 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


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

 15%|█▌        | 2/13 [00:10<00:58,  5.36s/it]

 23%|██▎       | 3/13 [00:17<00:58,  5.90s/it]

 31%|███       | 4/13 [00:28<01:10,  7.88s/it]

 38%|███▊      | 5/13 [00:40<01:13,  9.17s/it]

 46%|████▌     | 6/13 [00:51<01:09,  9.92s/it]

 54%|█████▍    | 7/13 [01:03<01:02, 10.46s/it]

 62%|██████▏   | 8/13 [01:14<00:53, 10.75s/it]

 69%|██████▉   | 9/13 [01:26<00:44, 11.04s/it]

 77%|███████▋  | 10/13 [01:37<00:33, 11.22s/it]

 85%|████████▍ | 11/13 [01:49<00:22, 11.30s/it]

 92%|█████████▏| 12/13 [02:00<00:11, 11.24s/it]

100%|██████████| 13/13 [02:11<00:00, 11.15s/it]2026-04-12 13:46:15,675 | INFO | train_grpo_train | metrics_logged mode=eval step=750

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 142.9322, 'eval_samples_per_second': 0.728, 'eval_steps_per_second': 0.091, 'eval_num_tokens': 1527911.0, 'eval_completions/mean_length': 246.6153846153846, 'eval_completions/min_length': 58.38461538461539, 'eval_completions/max_length': 482.2307692307692, 'eval_completions/clipped_ratio': 0.20192307692307693, 'eval_completions/mean_terminated_length': 179.07830106295071, 'eval_completions/min_terminated_length': 58.38461538461539, 'eval_completions/max_terminated_length': 331.6923076923077, 'eval_rewards/meter/mean': 0.5117177183811481, 'eval_rewards/meter/std': 0.36041144682810855, 'eval_rewards/count_adherence/mean': 0.8986238699692947, 'eval_rewards/count_adherence/std': 0.15372243007788292, 'eval_rewards/arabic_clean/mean': 0.9134615384615384, 'eval_rewards/arabic_clean/std': 0.20719900268774766, 'eval_rewards/lexical_plausibility/mean': 0.9704499336389395, 'eval_rewards/lexical_plausibility/std': 0.06162142452712242, 'eval_rewards/judge_quality/mean': 0.2686538489965292, 'eval_rewards/judge_quality/std': 0.14796888312468162, 'eval_rewards/repeat_penalty/mean': 0.5890747560904577, 'eval_rewards/repeat_penalty/std': 0.37633006274700165, 'eval_rewards/near_duplicate_penalty/mean': 0.854516694178948, 'eval_rewards/near_duplicate_penalty/std': 0.15290856862870547, 'eval_rewards/opening_diversity/mean': 0.9277158425404475, 'eval_rewards/opening_diversity/std': 0.14995075504367167, 'eval_rewards/distinct_2/mean': 0.6525185520832355, 'eval_rewards/distinct_2/std': 0.38338895715199983, 'eval_rewards/total_composite/mean': 0.11650820563618954, 'eval_rewards/total_composite/std': 0.10449972261603062, 'eval_reward': 0.11650820563618954, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.05291906085151892, 'eval_sampling/sampling_logp_difference/max': 1.1188103602482722, 'eval_sampling/importance_sampling_ratio/min': 0.3326175304559561, 'eval_sampling/importance_sampling_ratio/mean': 1.0090681589566743, 'eval_sampling/importance_sampling_ratio/max': 1.5178978076347938, 'eval_entropy': 0.5664004935668066, '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': 0.11650820563618954, 'eval_reward_meter_mean': 0.5117177183811481, 'eval_reward_meter_std': 0.36041144682810855, 'eval_reward_count_adherence_mean': 0.8986238699692947, 'eval_reward_count_adherence_std': 0.15372243007788292, 'eval_reward_arabic_clean_mean': 0.9134615384615384, 'eval_reward_arabic_clean_std': 0.20719900268774766, 'eval_reward_lexical_plausibility_mean': 0.9704499336389395, 'eval_reward_lexical_plausibility_std': 0.06162142452712242, 'eval_reward_repeat_penalty_mean': 0.5890747560904577, 'eval_reward_repeat_penalty_std': 0.37633006274700165, 'eval_reward_near_duplicate_penalty_mean': 0.854516694178948, 'eval_reward_near_duplicate_penalty_std': 0.15290856862870547, 'eval_reward_opening_diversity_mean': 0.9277158425404475, 'eval_reward_opening_diversity_std': 0.14995075504367167, 'eval_reward_distinct_2_mean': 0.6525185520832355, 'eval_reward_distinct_2_std': 0.38338895715199983, 'eval_reward_judge_quality_mean': 0.2686538489965292, 'eval_reward_judge_quality_std': 0.14796888312468162, 'eval_reward_total_composite_mean': 0.11650820563618954, 'eval_reward_total_composite_std': 0.10449972261603062, 'epoch': 0.03}

 23%|██▎       | 750/3300 [3:00:03<7:55:05, 11.18s/it]

100%|██████████| 13/13 [02:11<00:00, 11.15s/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 [3:00:20<39:26:07, 55.70s/it]2026-04-12 13:46:32,294 | INFO | train_grpo_train | metrics_logged mode=train step=751

                                                       
{'loss': -0.0016, 'grad_norm': 5.6542582511901855, 'learning_rate': 7.727272727272727e-06, 'num_tokens': 1529731.0, 'completions/mean_length': 115.5, 'completions/min_length': 48.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 58.857147216796875, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 88.0, 'rewards/meter/mean': 0.24534814059734344, 'rewards/meter/std': 0.3199789524078369, '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/lexical_plausibility/mean': 0.9051962494850159, 'rewards/lexical_plausibility/std': 0.14915074408054352, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.8770872950553894, 'rewards/repeat_penalty/std': 0.11442028731107712, 'rewards/near_duplicate_penalty/mean': 0.9448610544204712, 'rewards/near_duplicate_penalty/std': 0.03827836737036705, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9257956743240356, 'rewards/distinct_2/std': 0.09299763292074203, 'rewards/total_composite/mean': 0.06351770460605621, 'rewards/total_composite/std': 0.07437413930892944, 'reward': 0.06351770460605621, 'reward_std': 0.07437413930892944, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15945707261562347, 'sampling/sampling_logp_difference/max': 1.4806623458862305, 'sampling/importance_sampling_ratio/min': 0.2274869680404663, 'sampling/importance_sampling_ratio/mean': 1.005846619606018, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0602593272924423, 'clip_ratio/low_mean': 0.06553702801465988, 'clip_ratio/low_min': 0.06553702801465988, 'clip_ratio/high_mean': 0.06338721327483654, 'clip_ratio/high_max': 0.06338721327483654, 'clip_ratio/region_mean': 0.12892424128949642, 'reward_total_mean': 0.06351770460605621, 'reward_meter_mean': 0.24534814059734344, 'reward_meter_std': 0.3199789524078369, '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_lexical_plausibility_mean': 0.9051962494850159, 'reward_lexical_plausibility_std': 0.14915074408054352, 'reward_repeat_penalty_mean': 0.8770872950553894, 'reward_repeat_penalty_std': 0.11442028731107712, 'reward_near_duplicate_penalty_mean': 0.9448610544204712, 'reward_near_duplicate_penalty_std': 0.03827836737036705, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9257956743240356, 'reward_distinct_2_std': 0.09299763292074203, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.06351770460605621, 'reward_total_composite_std': 0.07437413930892944, 'epoch': 0.03}

 23%|██▎       | 751/3300 [3:00:20<39:26:07, 55.70s/it]INFO 04-12 13:46:32 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 752/3300 [3:00:34<30:34:14, 43.19s/it]2026-04-12 13:46:46,278 | INFO | train_grpo_train | metrics_logged mode=train step=752

                                                       
{'loss': -0.0446, 'grad_norm': 4.507869720458984, 'learning_rate': 7.724242424242424e-06, 'num_tokens': 1531328.0, 'completions/mean_length': 107.625, 'completions/min_length': 45.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 49.85714340209961, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.8025414943695068, 'rewards/meter/std': 0.3103002905845642, '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/lexical_plausibility/mean': 0.9397887587547302, 'rewards/lexical_plausibility/std': 0.13242894411087036, 'rewards/judge_quality/mean': 0.4124999940395355, 'rewards/judge_quality/std': 0.21339097619056702, 'rewards/repeat_penalty/mean': 0.8586175441741943, 'rewards/repeat_penalty/std': 0.1391318142414093, 'rewards/near_duplicate_penalty/mean': 0.921478271484375, 'rewards/near_duplicate_penalty/std': 0.040106385946273804, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.954170823097229, 'rewards/distinct_2/std': 0.054563283920288086, 'rewards/total_composite/mean': 0.2564723789691925, 'rewards/total_composite/std': 0.22441203892230988, 'reward': 0.2564723789691925, 'reward_std': 0.22441203892230988, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13170906901359558, 'sampling/sampling_logp_difference/max': 1.3046162128448486, 'sampling/importance_sampling_ratio/min': 0.32574662566185, 'sampling/importance_sampling_ratio/mean': 1.0027834177017212, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7087630107998848, 'clip_ratio/low_mean': 0.025000000838190317, 'clip_ratio/low_min': 0.025000000838190317, 'clip_ratio/high_mean': 0.08985042944550514, 'clip_ratio/high_max': 0.08985042944550514, 'clip_ratio/region_mean': 0.11485043028369546, 'reward_total_mean': 0.2564723789691925, 'reward_meter_mean': 0.8025414943695068, 'reward_meter_std': 0.3103002905845642, '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_lexical_plausibility_mean': 0.9397887587547302, 'reward_lexical_plausibility_std': 0.13242894411087036, 'reward_repeat_penalty_mean': 0.8586175441741943, 'reward_repeat_penalty_std': 0.1391318142414093, 'reward_near_duplicate_penalty_mean': 0.921478271484375, 'reward_near_duplicate_penalty_std': 0.040106385946273804, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.954170823097229, 'reward_distinct_2_std': 0.054563283920288086, 'reward_judge_quality_mean': 0.4124999940395355, 'reward_judge_quality_std': 0.21339097619056702, 'reward_total_composite_mean': 0.2564723789691925, 'reward_total_composite_std': 0.22441203892230988, 'epoch': 0.03}

 23%|██▎       | 752/3300 [3:00:34<30:34:14, 43.19s/it]INFO 04-12 13:46:46 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 753/3300 [3:00:47<24:09:19, 34.14s/it]2026-04-12 13:46:59,302 | INFO | train_grpo_train | metrics_logged mode=train step=753

                                                       
{'loss': -0.0272, 'grad_norm': 0.9113172292709351, 'learning_rate': 7.721212121212122e-06, 'num_tokens': 1532427.0, 'completions/mean_length': 265.375, 'completions/min_length': 17.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 18.75, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 21.0, 'rewards/meter/mean': 0.5618544816970825, 'rewards/meter/std': 0.4593985974788666, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/lexical_plausibility/mean': 0.7975413799285889, 'rewards/lexical_plausibility/std': 0.22083714604377747, 'rewards/judge_quality/mean': 0.4937499761581421, 'rewards/judge_quality/std': 0.4880262613296509, 'rewards/repeat_penalty/mean': 0.7169476747512817, 'rewards/repeat_penalty/std': 0.0651947557926178, 'rewards/near_duplicate_penalty/mean': 0.955930233001709, 'rewards/near_duplicate_penalty/std': 0.08692633360624313, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.35111892223358154, 'rewards/total_composite/std': 0.47708356380462646, 'reward': 0.35111892223358154, 'reward_std': 0.4770835340023041, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08600523322820663, 'sampling/sampling_logp_difference/max': 0.9025000333786011, 'sampling/importance_sampling_ratio/min': 0.43438971042633057, 'sampling/importance_sampling_ratio/mean': 1.0342092514038086, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.31766046583652496, 'clip_ratio/low_mean': 0.036764707416296005, 'clip_ratio/low_min': 0.036764707416296005, 'clip_ratio/high_mean': 0.012202381156384945, 'clip_ratio/high_max': 0.012202381156384945, 'clip_ratio/region_mean': 0.04896708857268095, 'reward_total_mean': 0.35111892223358154, 'reward_meter_mean': 0.5618544816970825, 'reward_meter_std': 0.4593985974788666, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_lexical_plausibility_mean': 0.7975413799285889, 'reward_lexical_plausibility_std': 0.22083714604377747, 'reward_repeat_penalty_mean': 0.7169476747512817, 'reward_repeat_penalty_std': 0.0651947557926178, 'reward_near_duplicate_penalty_mean': 0.955930233001709, 'reward_near_duplicate_penalty_std': 0.08692633360624313, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4937499761581421, 'reward_judge_quality_std': 0.4880262613296509, 'reward_total_composite_mean': 0.35111892223358154, 'reward_total_composite_std': 0.47708356380462646, 'epoch': 0.03}

 23%|██▎       | 753/3300 [3:00:47<24:09:19, 34.14s/it]INFO 04-12 13:46:59 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 754/3300 [3:00:59<19:35:27, 27.70s/it]2026-04-12 13:47:11,973 | INFO | train_grpo_train | metrics_logged mode=train step=754

                                                       
{'loss': -0.0599, 'grad_norm': 4.117982864379883, 'learning_rate': 7.718181818181819e-06, 'num_tokens': 1533982.0, 'completions/mean_length': 93.375, 'completions/min_length': 28.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 33.57143020629883, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.8072152137756348, 'rewards/meter/std': 0.33680295944213867, '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/lexical_plausibility/mean': 0.8812618255615234, 'rewards/lexical_plausibility/std': 0.20129022002220154, 'rewards/judge_quality/mean': 0.6312500238418579, 'rewards/judge_quality/std': 0.4020816683769226, 'rewards/repeat_penalty/mean': 0.7487483024597168, 'rewards/repeat_penalty/std': 0.1261836141347885, 'rewards/near_duplicate_penalty/mean': 0.9697259664535522, 'rewards/near_duplicate_penalty/std': 0.06563077121973038, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9839743375778198, 'rewards/distinct_2/std': 0.045327357947826385, 'rewards/total_composite/mean': 0.5825003981590271, 'rewards/total_composite/std': 0.39257144927978516, 'reward': 0.5825003981590271, 'reward_std': 0.39257141947746277, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1541319191455841, 'sampling/sampling_logp_difference/max': 1.615778923034668, 'sampling/importance_sampling_ratio/min': 0.1987358182668686, 'sampling/importance_sampling_ratio/mean': 1.0031853914260864, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8037475720047951, 'clip_ratio/low_mean': 0.01923963101580739, 'clip_ratio/low_min': 0.01923963101580739, 'clip_ratio/high_mean': 0.08317217603325844, 'clip_ratio/high_max': 0.08317217603325844, 'clip_ratio/region_mean': 0.10241180704906583, 'reward_total_mean': 0.5825003981590271, 'reward_meter_mean': 0.8072152137756348, 'reward_meter_std': 0.33680295944213867, '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_lexical_plausibility_mean': 0.8812618255615234, 'reward_lexical_plausibility_std': 0.20129022002220154, 'reward_repeat_penalty_mean': 0.7487483024597168, 'reward_repeat_penalty_std': 0.1261836141347885, 'reward_near_duplicate_penalty_mean': 0.9697259664535522, 'reward_near_duplicate_penalty_std': 0.06563077121973038, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9839743375778198, 'reward_distinct_2_std': 0.045327357947826385, 'reward_judge_quality_mean': 0.6312500238418579, 'reward_judge_quality_std': 0.4020816683769226, 'reward_total_composite_mean': 0.5825003981590271, 'reward_total_composite_std': 0.39257144927978516, 'epoch': 0.03}

 23%|██▎       | 754/3300 [3:00:59<19:35:27, 27.70s/it]INFO 04-12 13:47:12 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 755/3300 [3:01:07<15:25:31, 21.82s/it]2026-04-12 13:47:20,069 | INFO | train_grpo_train | metrics_logged mode=train step=755

                                                       
{'loss': 0.0474, 'grad_norm': 14.829696655273438, 'learning_rate': 7.715151515151516e-06, 'num_tokens': 1535459.0, 'completions/mean_length': 33.625, 'completions/min_length': 30.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.625, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.7113251686096191, 'rewards/meter/std': 0.386238694190979, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9771242141723633, 'rewards/lexical_plausibility/std': 0.0647025778889656, 'rewards/judge_quality/mean': 0.5637500286102295, 'rewards/judge_quality/std': 0.3091896176338196, 'rewards/repeat_penalty/mean': 0.7080320119857788, 'rewards/repeat_penalty/std': 0.0597260519862175, 'rewards/near_duplicate_penalty/mean': 0.9440426826477051, 'rewards/near_duplicate_penalty/std': 0.07963474839925766, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.45766425132751465, 'rewards/total_composite/std': 0.3889073133468628, 'reward': 0.45766425132751465, 'reward_std': 0.3889073133468628, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15598982572555542, 'sampling/sampling_logp_difference/max': 2.367280960083008, 'sampling/importance_sampling_ratio/min': 0.09373525530099869, 'sampling/importance_sampling_ratio/mean': 1.0085208415985107, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0116269662976265, 'clip_ratio/low_mean': 0.07379679288715124, 'clip_ratio/low_min': 0.07379679288715124, 'clip_ratio/high_mean': 0.07753136195242405, 'clip_ratio/high_max': 0.07753136195242405, 'clip_ratio/region_mean': 0.1513281548395753, 'reward_total_mean': 0.45766425132751465, 'reward_meter_mean': 0.7113251686096191, 'reward_meter_std': 0.386238694190979, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9771242141723633, 'reward_lexical_plausibility_std': 0.0647025778889656, 'reward_repeat_penalty_mean': 0.7080320119857788, 'reward_repeat_penalty_std': 0.0597260519862175, 'reward_near_duplicate_penalty_mean': 0.9440426826477051, 'reward_near_duplicate_penalty_std': 0.07963474839925766, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5637500286102295, 'reward_judge_quality_std': 0.3091896176338196, 'reward_total_composite_mean': 0.45766425132751465, 'reward_total_composite_std': 0.3889073133468628, 'epoch': 0.03}

 23%|██▎       | 755/3300 [3:01:08<15:25:31, 21.82s/it]INFO 04-12 13:47:20 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 756/3300 [3:01:21<13:41:40, 19.38s/it]2026-04-12 13:47:33,758 | INFO | train_grpo_train | metrics_logged mode=train step=756

                                                       
{'loss': -0.0798, 'grad_norm': 5.820659637451172, 'learning_rate': 7.712121212121213e-06, 'num_tokens': 1537325.0, 'completions/mean_length': 128.25, 'completions/min_length': 61.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 73.42857360839844, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.7581970691680908, 'rewards/meter/std': 0.2884212136268616, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9799602627754211, 'rewards/lexical_plausibility/std': 0.05668089911341667, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.7422148585319519, 'rewards/repeat_penalty/std': 0.28219759464263916, 'rewards/near_duplicate_penalty/mean': 0.9217819571495056, 'rewards/near_duplicate_penalty/std': 0.10180716961622238, 'rewards/opening_diversity/mean': 0.887499988079071, 'rewards/opening_diversity/std': 0.21001701056957245, 'rewards/distinct_2/mean': 0.8897054195404053, 'rewards/distinct_2/std': 0.1511707305908203, 'rewards/total_composite/mean': 0.20604243874549866, 'rewards/total_composite/std': 0.1265801638364792, 'reward': 0.20604243874549866, 'reward_std': 0.1265801638364792, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.139238178730011, 'sampling/sampling_logp_difference/max': 1.6953731775283813, 'sampling/importance_sampling_ratio/min': 0.18353073298931122, 'sampling/importance_sampling_ratio/mean': 1.0003151893615723, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7519563883543015, 'clip_ratio/low_mean': 0.032273298129439354, 'clip_ratio/low_min': 0.032273298129439354, 'clip_ratio/high_mean': 0.06528119184076786, 'clip_ratio/high_max': 0.06528119184076786, 'clip_ratio/region_mean': 0.09755448997020721, 'reward_total_mean': 0.20604243874549866, 'reward_meter_mean': 0.7581970691680908, 'reward_meter_std': 0.2884212136268616, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9799602627754211, 'reward_lexical_plausibility_std': 0.05668089911341667, 'reward_repeat_penalty_mean': 0.7422148585319519, 'reward_repeat_penalty_std': 0.28219759464263916, 'reward_near_duplicate_penalty_mean': 0.9217819571495056, 'reward_near_duplicate_penalty_std': 0.10180716961622238, 'reward_opening_diversity_mean': 0.887499988079071, 'reward_opening_diversity_std': 0.21001701056957245, 'reward_distinct_2_mean': 0.8897054195404053, 'reward_distinct_2_std': 0.1511707305908203, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.20604243874549866, 'reward_total_composite_std': 0.1265801638364792, 'epoch': 0.03}

 23%|██▎       | 756/3300 [3:01:21<13:41:40, 19.38s/it]INFO 04-12 13:47:34 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 757/3300 [3:01:32<11:50:47, 16.77s/it]2026-04-12 13:47:44,438 | INFO | train_grpo_train | metrics_logged mode=train step=757

                                                       
{'loss': 0.0844, 'grad_norm': 6.699467182159424, 'learning_rate': 7.709090909090909e-06, 'num_tokens': 1540312.0, 'completions/mean_length': 180.375, 'completions/min_length': 150.0, 'completions/max_length': 210.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 180.375, 'completions/min_terminated_length': 150.0, 'completions/max_terminated_length': 210.0, 'rewards/meter/mean': 0.9924758672714233, 'rewards/meter/std': 0.004042538348585367, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.995192289352417, 'rewards/lexical_plausibility/std': 0.013598216697573662, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.23643825948238373, 'rewards/repeat_penalty/std': 0.1517748236656189, 'rewards/near_duplicate_penalty/mean': 0.7452784776687622, 'rewards/near_duplicate_penalty/std': 0.09216246008872986, 'rewards/opening_diversity/mean': 0.878125011920929, 'rewards/opening_diversity/std': 0.2217775136232376, 'rewards/distinct_2/mean': 0.3658512234687805, 'rewards/distinct_2/std': 0.14827655255794525, 'rewards/total_composite/mean': 0.14143750071525574, 'rewards/total_composite/std': 0.013895310461521149, 'reward': 0.14143750071525574, 'reward_std': 0.013895308598876, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07771893590688705, 'sampling/sampling_logp_difference/max': 1.7559001445770264, 'sampling/importance_sampling_ratio/min': 0.17275166511535645, 'sampling/importance_sampling_ratio/mean': 1.0070961713790894, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.555378720164299, 'clip_ratio/low_mean': 0.018682016525417566, 'clip_ratio/low_min': 0.018682016525417566, 'clip_ratio/high_mean': 0.06278140377253294, 'clip_ratio/high_max': 0.06278140377253294, 'clip_ratio/region_mean': 0.0814634202979505, 'reward_total_mean': 0.14143750071525574, 'reward_meter_mean': 0.9924758672714233, 'reward_meter_std': 0.004042538348585367, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.09258200973272324, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.995192289352417, 'reward_lexical_plausibility_std': 0.013598216697573662, 'reward_repeat_penalty_mean': 0.23643825948238373, 'reward_repeat_penalty_std': 0.1517748236656189, 'reward_near_duplicate_penalty_mean': 0.7452784776687622, 'reward_near_duplicate_penalty_std': 0.09216246008872986, 'reward_opening_diversity_mean': 0.878125011920929, 'reward_opening_diversity_std': 0.2217775136232376, 'reward_distinct_2_mean': 0.3658512234687805, 'reward_distinct_2_std': 0.14827655255794525, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.14143750071525574, 'reward_total_composite_std': 0.013895310461521149, 'epoch': 0.03}

 23%|██▎       | 757/3300 [3:01:32<11:50:47, 16.77s/it]INFO 04-12 13:47:44 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 758/3300 [3:01:43<10:43:04, 15.18s/it]2026-04-12 13:47:55,908 | INFO | train_grpo_train | metrics_logged mode=train step=758

                                                       
{'loss': -0.012, 'grad_norm': 2.2713780403137207, 'learning_rate': 7.706060606060606e-06, 'num_tokens': 1544597.0, 'completions/mean_length': 317.625, 'completions/min_length': 295.0, 'completions/max_length': 343.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 317.625, 'completions/min_terminated_length': 295.0, 'completions/max_terminated_length': 343.0, 'rewards/meter/mean': 0.7525134682655334, 'rewards/meter/std': 0.4262067377567291, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.025922738015651703, 'rewards/repeat_penalty/std': 0.044997766613960266, 'rewards/near_duplicate_penalty/mean': 0.5169505476951599, 'rewards/near_duplicate_penalty/std': 0.1723085194826126, 'rewards/opening_diversity/mean': 0.46875, 'rewards/opening_diversity/std': 0.29859957098960876, 'rewards/distinct_2/mean': 0.07271120697259903, 'rewards/distinct_2/std': 0.13307280838489532, 'rewards/total_composite/mean': 0.10095322877168655, 'rewards/total_composite/std': 0.05731987953186035, 'reward': 0.10095322877168655, 'reward_std': 0.05731987580657005, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04793323203921318, 'sampling/sampling_logp_difference/max': 2.0491909980773926, 'sampling/importance_sampling_ratio/min': 0.12883909046649933, 'sampling/importance_sampling_ratio/mean': 1.0077365636825562, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2946852520108223, 'clip_ratio/low_mean': 0.008336049737408757, 'clip_ratio/low_min': 0.008336049737408757, 'clip_ratio/high_mean': 0.047590904869139194, 'clip_ratio/high_max': 0.047590904869139194, 'clip_ratio/region_mean': 0.05592695460654795, 'reward_total_mean': 0.10095322877168655, 'reward_meter_mean': 0.7525134682655334, 'reward_meter_std': 0.4262067377567291, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.025922738015651703, 'reward_repeat_penalty_std': 0.044997766613960266, 'reward_near_duplicate_penalty_mean': 0.5169505476951599, 'reward_near_duplicate_penalty_std': 0.1723085194826126, 'reward_opening_diversity_mean': 0.46875, 'reward_opening_diversity_std': 0.29859957098960876, 'reward_distinct_2_mean': 0.07271120697259903, 'reward_distinct_2_std': 0.13307280838489532, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.10095322877168655, 'reward_total_composite_std': 0.05731987953186035, 'epoch': 0.03}

 23%|██▎       | 758/3300 [3:01:43<10:43:04, 15.18s/it]INFO 04-12 13:47:56 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 759/3300 [3:01:57<10:26:45, 14.80s/it]2026-04-12 13:48:09,822 | INFO | train_grpo_train | metrics_logged mode=train step=759

                                                       
{'loss': -0.0843, 'grad_norm': 5.156977653503418, 'learning_rate': 7.703030303030304e-06, 'num_tokens': 1546478.0, 'completions/mean_length': 132.125, 'completions/min_length': 68.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 77.85714721679688, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 85.0, 'rewards/meter/mean': 0.617965579032898, 'rewards/meter/std': 0.3356782793998718, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.2357022762298584, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9563086032867432, 'rewards/lexical_plausibility/std': 0.12357787042856216, 'rewards/judge_quality/mean': 0.36249998211860657, 'rewards/judge_quality/std': 0.13562026619911194, 'rewards/repeat_penalty/mean': 0.8897678256034851, 'rewards/repeat_penalty/std': 0.09109393507242203, 'rewards/near_duplicate_penalty/mean': 0.9475877285003662, 'rewards/near_duplicate_penalty/std': 0.04304119572043419, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9704861044883728, 'rewards/distinct_2/std': 0.04337109625339508, 'rewards/total_composite/mean': 0.1589117795228958, 'rewards/total_composite/std': 0.1443527787923813, 'reward': 0.1589117795228958, 'reward_std': 0.1443527638912201, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12779755890369415, 'sampling/sampling_logp_difference/max': 1.3012678623199463, 'sampling/importance_sampling_ratio/min': 0.2721864879131317, 'sampling/importance_sampling_ratio/mean': 1.0205940008163452, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7309146523475647, 'clip_ratio/low_mean': 0.04924075026065111, 'clip_ratio/low_min': 0.04924075026065111, 'clip_ratio/high_mean': 0.045848920941352844, 'clip_ratio/high_max': 0.045848920941352844, 'clip_ratio/region_mean': 0.09508967120200396, 'reward_total_mean': 0.1589117795228958, 'reward_meter_mean': 0.617965579032898, 'reward_meter_std': 0.3356782793998718, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.2357022762298584, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9563086032867432, 'reward_lexical_plausibility_std': 0.12357787042856216, 'reward_repeat_penalty_mean': 0.8897678256034851, 'reward_repeat_penalty_std': 0.09109393507242203, 'reward_near_duplicate_penalty_mean': 0.9475877285003662, 'reward_near_duplicate_penalty_std': 0.04304119572043419, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9704861044883728, 'reward_distinct_2_std': 0.04337109625339508, 'reward_judge_quality_mean': 0.36249998211860657, 'reward_judge_quality_std': 0.13562026619911194, 'reward_total_composite_mean': 0.1589117795228958, 'reward_total_composite_std': 0.1443527787923813, 'epoch': 0.03}

 23%|██▎       | 759/3300 [3:01:57<10:26:45, 14.80s/it]INFO 04-12 13:48:10 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 760/3300 [3:02:10<10:05:36, 14.31s/it]2026-04-12 13:48:22,972 | INFO | train_grpo_train | metrics_logged mode=train step=760

                                                       
{'loss': -0.03, 'grad_norm': 7.912317752838135, 'learning_rate': 7.7e-06, 'num_tokens': 1548105.0, 'completions/mean_length': 96.375, 'completions/min_length': 35.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 37.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.6226221323013306, 'rewards/meter/std': 0.3957872688770294, '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/lexical_plausibility/mean': 0.9024482369422913, 'rewards/lexical_plausibility/std': 0.18081353604793549, 'rewards/judge_quality/mean': 0.4424999952316284, 'rewards/judge_quality/std': 0.32101401686668396, 'rewards/repeat_penalty/mean': 0.9369135499000549, 'rewards/repeat_penalty/std': 0.09812384843826294, 'rewards/near_duplicate_penalty/mean': 0.9622968435287476, 'rewards/near_duplicate_penalty/std': 0.05941963568329811, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.971470832824707, 'rewards/distinct_2/std': 0.05287402868270874, 'rewards/total_composite/mean': 0.2968449592590332, 'rewards/total_composite/std': 0.2879330813884735, 'reward': 0.2968449592590332, 'reward_std': 0.2879330813884735, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14481541514396667, 'sampling/sampling_logp_difference/max': 1.2011079788208008, 'sampling/importance_sampling_ratio/min': 0.3008606731891632, 'sampling/importance_sampling_ratio/mean': 1.0089575052261353, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9489739835262299, 'clip_ratio/low_mean': 0.029910714365541935, 'clip_ratio/low_min': 0.029910714365541935, 'clip_ratio/high_mean': 0.09513889090158045, 'clip_ratio/high_max': 0.09513889090158045, 'clip_ratio/region_mean': 0.1250496052671224, 'reward_total_mean': 0.2968449592590332, 'reward_meter_mean': 0.6226221323013306, 'reward_meter_std': 0.3957872688770294, '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_lexical_plausibility_mean': 0.9024482369422913, 'reward_lexical_plausibility_std': 0.18081353604793549, 'reward_repeat_penalty_mean': 0.9369135499000549, 'reward_repeat_penalty_std': 0.09812384843826294, 'reward_near_duplicate_penalty_mean': 0.9622968435287476, 'reward_near_duplicate_penalty_std': 0.05941963568329811, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.971470832824707, 'reward_distinct_2_std': 0.05287402868270874, 'reward_judge_quality_mean': 0.4424999952316284, 'reward_judge_quality_std': 0.32101401686668396, 'reward_total_composite_mean': 0.2968449592590332, 'reward_total_composite_std': 0.2879330813884735, 'epoch': 0.03}

 23%|██▎       | 760/3300 [3:02:10<10:05:36, 14.31s/it]INFO 04-12 13:48:23 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 761/3300 [3:02:24<9:54:21, 14.05s/it] 2026-04-12 13:48:36,412 | INFO | train_grpo_train | metrics_logged mode=train step=761

                                                      
{'loss': -0.0299, 'grad_norm': 5.511273384094238, 'learning_rate': 7.696969696969696e-06, 'num_tokens': 1549770.0, 'completions/mean_length': 105.125, 'completions/min_length': 44.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 47.000003814697266, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.4457665979862213, 'rewards/meter/std': 0.4284104108810425, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.959851861000061, 'rewards/lexical_plausibility/std': 0.11355604976415634, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.16035674512386322, 'rewards/repeat_penalty/mean': 0.9313533306121826, 'rewards/repeat_penalty/std': 0.10008540004491806, 'rewards/near_duplicate_penalty/mean': 0.9754289984703064, 'rewards/near_duplicate_penalty/std': 0.021154019981622696, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9536713361740112, 'rewards/distinct_2/std': 0.08915838599205017, 'rewards/total_composite/mean': 0.18814267218112946, 'rewards/total_composite/std': 0.17935602366924286, 'reward': 0.18814267218112946, 'reward_std': 0.17935602366924286, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15126250684261322, 'sampling/sampling_logp_difference/max': 2.151500701904297, 'sampling/importance_sampling_ratio/min': 0.11630947887897491, 'sampling/importance_sampling_ratio/mean': 1.0220063924789429, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8205545321106911, 'clip_ratio/low_mean': 0.04237702256068587, 'clip_ratio/low_min': 0.04237702256068587, 'clip_ratio/high_mean': 0.07541702594608068, 'clip_ratio/high_max': 0.07541702594608068, 'clip_ratio/region_mean': 0.11779404850676656, 'reward_total_mean': 0.18814267218112946, 'reward_meter_mean': 0.4457665979862213, 'reward_meter_std': 0.4284104108810425, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.959851861000061, 'reward_lexical_plausibility_std': 0.11355604976415634, 'reward_repeat_penalty_mean': 0.9313533306121826, 'reward_repeat_penalty_std': 0.10008540004491806, 'reward_near_duplicate_penalty_mean': 0.9754289984703064, 'reward_near_duplicate_penalty_std': 0.021154019981622696, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9536713361740112, 'reward_distinct_2_std': 0.08915838599205017, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.16035674512386322, 'reward_total_composite_mean': 0.18814267218112946, 'reward_total_composite_std': 0.17935602366924286, 'epoch': 0.03}

 23%|██▎       | 761/3300 [3:02:24<9:54:21, 14.05s/it]INFO 04-12 13:48:36 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 762/3300 [3:02:39<10:03:24, 14.27s/it]2026-04-12 13:48:51,186 | INFO | train_grpo_train | metrics_logged mode=train step=762

                                                       
{'loss': -0.0326, 'grad_norm': 5.418692588806152, 'learning_rate': 7.693939393939395e-06, 'num_tokens': 1551418.0, 'completions/mean_length': 115.0, 'completions/min_length': 50.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 58.28571701049805, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.26285335421562195, 'rewards/meter/std': 0.26476600766181946, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.24800792336463928, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9494400024414062, 'rewards/lexical_plausibility/std': 0.14300529658794403, 'rewards/judge_quality/mean': 0.4399999976158142, 'rewards/judge_quality/std': 0.24859607219696045, 'rewards/repeat_penalty/mean': 0.9500300884246826, 'rewards/repeat_penalty/std': 0.08432653546333313, 'rewards/near_duplicate_penalty/mean': 0.9895552396774292, 'rewards/near_duplicate_penalty/std': 0.009525802917778492, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9916387796401978, 'rewards/distinct_2/std': 0.0236490610986948, 'rewards/total_composite/mean': 0.09958121180534363, 'rewards/total_composite/std': 0.1242804005742073, 'reward': 0.09958121180534363, 'reward_std': 0.1242804005742073, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1265241503715515, 'sampling/sampling_logp_difference/max': 1.0849881172180176, 'sampling/importance_sampling_ratio/min': 0.3379058241844177, 'sampling/importance_sampling_ratio/mean': 1.0120832920074463, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7335836812853813, 'clip_ratio/low_mean': 0.06432131212204695, 'clip_ratio/low_min': 0.06432131212204695, 'clip_ratio/high_mean': 0.05025284644216299, 'clip_ratio/high_max': 0.05025284644216299, 'clip_ratio/region_mean': 0.11457415856420994, 'reward_total_mean': 0.09958121180534363, 'reward_meter_mean': 0.26285335421562195, 'reward_meter_std': 0.26476600766181946, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.24800792336463928, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9494400024414062, 'reward_lexical_plausibility_std': 0.14300529658794403, 'reward_repeat_penalty_mean': 0.9500300884246826, 'reward_repeat_penalty_std': 0.08432653546333313, 'reward_near_duplicate_penalty_mean': 0.9895552396774292, 'reward_near_duplicate_penalty_std': 0.009525802917778492, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9916387796401978, 'reward_distinct_2_std': 0.0236490610986948, 'reward_judge_quality_mean': 0.4399999976158142, 'reward_judge_quality_std': 0.24859607219696045, 'reward_total_composite_mean': 0.09958121180534363, 'reward_total_composite_std': 0.1242804005742073, 'epoch': 0.03}

 23%|██▎       | 762/3300 [3:02:39<10:03:24, 14.27s/it]INFO 04-12 13:48:51 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 763/3300 [3:02:50<9:23:21, 13.32s/it] 2026-04-12 13:49:02,312 | INFO | train_grpo_train | metrics_logged mode=train step=763

                                                      
{'loss': 0.0566, 'grad_norm': 4.991977214813232, 'learning_rate': 7.690909090909091e-06, 'num_tokens': 1554835.0, 'completions/mean_length': 224.125, 'completions/min_length': 199.0, 'completions/max_length': 246.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 224.125, 'completions/min_terminated_length': 199.0, 'completions/max_terminated_length': 246.0, 'rewards/meter/mean': 0.7947300672531128, 'rewards/meter/std': 0.28722473978996277, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.2018357515335083, 'rewards/repeat_penalty/std': 0.14820127189159393, 'rewards/near_duplicate_penalty/mean': 0.7426934242248535, 'rewards/near_duplicate_penalty/std': 0.11225884407758713, 'rewards/opening_diversity/mean': 0.9553571939468384, 'rewards/opening_diversity/std': 0.06259103864431381, 'rewards/distinct_2/mean': 0.3077459931373596, 'rewards/distinct_2/std': 0.21564757823944092, 'rewards/total_composite/mean': 0.11920950561761856, 'rewards/total_composite/std': 0.043083708733320236, 'reward': 0.11920950561761856, 'reward_std': 0.043083708733320236, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08441130071878433, 'sampling/sampling_logp_difference/max': 1.802474021911621, 'sampling/importance_sampling_ratio/min': 0.16489045321941376, 'sampling/importance_sampling_ratio/mean': 1.0006810426712036, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.46828923374414444, 'clip_ratio/low_mean': 0.024803034961223602, 'clip_ratio/low_min': 0.024803034961223602, 'clip_ratio/high_mean': 0.059402203653007746, 'clip_ratio/high_max': 0.059402203653007746, 'clip_ratio/region_mean': 0.08420523861423135, 'reward_total_mean': 0.11920950561761856, 'reward_meter_mean': 0.7947300672531128, 'reward_meter_std': 0.28722473978996277, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.2018357515335083, 'reward_repeat_penalty_std': 0.14820127189159393, 'reward_near_duplicate_penalty_mean': 0.7426934242248535, 'reward_near_duplicate_penalty_std': 0.11225884407758713, 'reward_opening_diversity_mean': 0.9553571939468384, 'reward_opening_diversity_std': 0.06259103864431381, 'reward_distinct_2_mean': 0.3077459931373596, 'reward_distinct_2_std': 0.21564757823944092, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.11920950561761856, 'reward_total_composite_std': 0.043083708733320236, 'epoch': 0.03}

 23%|██▎       | 763/3300 [3:02:50<9:23:21, 13.32s/it]INFO 04-12 13:49:02 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 764/3300 [3:03:04<9:35:57, 13.63s/it]2026-04-12 13:49:16,649 | INFO | train_grpo_train | metrics_logged mode=train step=764

                                                      
{'loss': -0.0992, 'grad_norm': 4.293631553649902, 'learning_rate': 7.687878787878788e-06, 'num_tokens': 1556515.0, 'completions/mean_length': 113.0, 'completions/min_length': 50.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 56.000003814697266, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.7618042826652527, 'rewards/meter/std': 0.3373679518699646, '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/lexical_plausibility/mean': 0.9540361762046814, 'rewards/lexical_plausibility/std': 0.13000528514385223, 'rewards/judge_quality/mean': 0.4337499737739563, 'rewards/judge_quality/std': 0.23868314921855927, 'rewards/repeat_penalty/mean': 0.9198135137557983, 'rewards/repeat_penalty/std': 0.11262106895446777, 'rewards/near_duplicate_penalty/mean': 0.9657071828842163, 'rewards/near_duplicate_penalty/std': 0.027733059599995613, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9807209968566895, 'rewards/distinct_2/std': 0.0357571542263031, 'rewards/total_composite/mean': 0.34798580408096313, 'rewards/total_composite/std': 0.2204636186361313, 'reward': 0.34798580408096313, 'reward_std': 0.2204636037349701, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14719754457473755, 'sampling/sampling_logp_difference/max': 1.2056770324707031, 'sampling/importance_sampling_ratio/min': 0.29948917031288147, 'sampling/importance_sampling_ratio/mean': 0.9965231418609619, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8594876527786255, 'clip_ratio/low_mean': 0.029282765928655863, 'clip_ratio/low_min': 0.029282765928655863, 'clip_ratio/high_mean': 0.09677751548588276, 'clip_ratio/high_max': 0.09677751548588276, 'clip_ratio/region_mean': 0.12606028141453862, 'reward_total_mean': 0.34798580408096313, 'reward_meter_mean': 0.7618042826652527, 'reward_meter_std': 0.3373679518699646, '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_lexical_plausibility_mean': 0.9540361762046814, 'reward_lexical_plausibility_std': 0.13000528514385223, 'reward_repeat_penalty_mean': 0.9198135137557983, 'reward_repeat_penalty_std': 0.11262106895446777, 'reward_near_duplicate_penalty_mean': 0.9657071828842163, 'reward_near_duplicate_penalty_std': 0.027733059599995613, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9807209968566895, 'reward_distinct_2_std': 0.0357571542263031, 'reward_judge_quality_mean': 0.4337499737739563, 'reward_judge_quality_std': 0.23868314921855927, 'reward_total_composite_mean': 0.34798580408096313, 'reward_total_composite_std': 0.2204636186361313, 'epoch': 0.03}

 23%|██▎       | 764/3300 [3:03:04<9:35:57, 13.63s/it]INFO 04-12 13:49:16 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 765/3300 [3:03:18<9:42:56, 13.80s/it]2026-04-12 13:49:30,843 | INFO | train_grpo_train | metrics_logged mode=train step=765

                                                      
{'loss': -0.1149, 'grad_norm': 2.4354758262634277, 'learning_rate': 7.684848484848485e-06, 'num_tokens': 1558242.0, 'completions/mean_length': 243.875, 'completions/min_length': 75.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 83.0, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 91.0, 'rewards/meter/mean': 0.2985907196998596, 'rewards/meter/std': 0.3216824531555176, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.8602477312088013, 'rewards/lexical_plausibility/std': 0.1957496851682663, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.19086270034313202, 'rewards/repeat_penalty/mean': 0.9494669437408447, 'rewards/repeat_penalty/std': 0.031903624534606934, 'rewards/near_duplicate_penalty/mean': 0.9611002802848816, 'rewards/near_duplicate_penalty/std': 0.02690417692065239, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.987949550151825, 'rewards/distinct_2/std': 0.02104729413986206, 'rewards/total_composite/mean': 0.05845516175031662, 'rewards/total_composite/std': 0.06506701558828354, 'reward': 0.05845516175031662, 'reward_std': 0.06506701558828354, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11565515398979187, 'sampling/sampling_logp_difference/max': 2.8310375213623047, 'sampling/importance_sampling_ratio/min': 0.05895165726542473, 'sampling/importance_sampling_ratio/mean': 1.0137100219726562, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5242686420679092, 'clip_ratio/low_mean': 0.01406249962747097, 'clip_ratio/low_min': 0.01406249962747097, 'clip_ratio/high_mean': 0.05406558606773615, 'clip_ratio/high_max': 0.05406558606773615, 'clip_ratio/region_mean': 0.06812808569520712, 'reward_total_mean': 0.05845516175031662, 'reward_meter_mean': 0.2985907196998596, 'reward_meter_std': 0.3216824531555176, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.8602477312088013, 'reward_lexical_plausibility_std': 0.1957496851682663, 'reward_repeat_penalty_mean': 0.9494669437408447, 'reward_repeat_penalty_std': 0.031903624534606934, 'reward_near_duplicate_penalty_mean': 0.9611002802848816, 'reward_near_duplicate_penalty_std': 0.02690417692065239, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.987949550151825, 'reward_distinct_2_std': 0.02104729413986206, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.19086270034313202, 'reward_total_composite_mean': 0.05845516175031662, 'reward_total_composite_std': 0.06506701558828354, 'epoch': 0.03}

 23%|██▎       | 765/3300 [3:03:18<9:42:56, 13.80s/it]INFO 04-12 13:49:31 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 766/3300 [3:03:33<9:50:00, 13.97s/it]2026-04-12 13:49:45,219 | INFO | train_grpo_train | metrics_logged mode=train step=766

                                                      
{'loss': -0.0272, 'grad_norm': 4.402597427368164, 'learning_rate': 7.681818181818183e-06, 'num_tokens': 1559787.0, 'completions/mean_length': 98.125, 'completions/min_length': 35.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 39.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.6116056442260742, 'rewards/meter/std': 0.36475294828414917, '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/lexical_plausibility/mean': 0.955929160118103, 'rewards/lexical_plausibility/std': 0.12465112656354904, 'rewards/judge_quality/mean': 0.4675000309944153, 'rewards/judge_quality/std': 0.3022179901599884, 'rewards/repeat_penalty/mean': 0.9904035329818726, 'rewards/repeat_penalty/std': 0.010543355718255043, 'rewards/near_duplicate_penalty/mean': 0.9904035329818726, 'rewards/near_duplicate_penalty/std': 0.010543355718255043, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.28405898809432983, 'rewards/total_composite/std': 0.2783108353614807, 'reward': 0.28405898809432983, 'reward_std': 0.2783108353614807, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11154511570930481, 'sampling/sampling_logp_difference/max': 2.150965452194214, 'sampling/importance_sampling_ratio/min': 0.116371750831604, 'sampling/importance_sampling_ratio/mean': 1.0186054706573486, 'sampling/importance_sampling_ratio/max': 1.734793782234192, 'entropy': 0.6226985715329647, 'clip_ratio/low_mean': 0.06483082100749016, 'clip_ratio/low_min': 0.06483082100749016, 'clip_ratio/high_mean': 0.06758247036486864, 'clip_ratio/high_max': 0.06758247036486864, 'clip_ratio/region_mean': 0.1324132913723588, 'reward_total_mean': 0.28405898809432983, 'reward_meter_mean': 0.6116056442260742, 'reward_meter_std': 0.36475294828414917, '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_lexical_plausibility_mean': 0.955929160118103, 'reward_lexical_plausibility_std': 0.12465112656354904, 'reward_repeat_penalty_mean': 0.9904035329818726, 'reward_repeat_penalty_std': 0.010543355718255043, 'reward_near_duplicate_penalty_mean': 0.9904035329818726, 'reward_near_duplicate_penalty_std': 0.010543355718255043, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4675000309944153, 'reward_judge_quality_std': 0.3022179901599884, 'reward_total_composite_mean': 0.28405898809432983, 'reward_total_composite_std': 0.2783108353614807, 'epoch': 0.03}

 23%|██▎       | 766/3300 [3:03:33<9:50:00, 13.97s/it]INFO 04-12 13:49:45 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 767/3300 [3:03:46<9:46:58, 13.90s/it]2026-04-12 13:49:58,967 | INFO | train_grpo_train | metrics_logged mode=train step=767

                                                      
{'loss': -0.0179, 'grad_norm': 4.854854106903076, 'learning_rate': 7.678787878787878e-06, 'num_tokens': 1561400.0, 'completions/mean_length': 109.625, 'completions/min_length': 43.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 52.142860412597656, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.24479930102825165, 'rewards/meter/std': 0.3385663628578186, '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/lexical_plausibility/mean': 0.9592169523239136, 'rewards/lexical_plausibility/std': 0.10458080470561981, 'rewards/judge_quality/mean': 0.3712499737739563, 'rewards/judge_quality/std': 0.25670647621154785, 'rewards/repeat_penalty/mean': 0.8317855000495911, 'rewards/repeat_penalty/std': 0.17168274521827698, 'rewards/near_duplicate_penalty/mean': 0.9257292747497559, 'rewards/near_duplicate_penalty/std': 0.04465325176715851, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9487742185592651, 'rewards/distinct_2/std': 0.07594232261180878, 'rewards/total_composite/mean': 0.15713383257389069, 'rewards/total_composite/std': 0.29908961057662964, 'reward': 0.15713383257389069, 'reward_std': 0.29908961057662964, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16317471861839294, 'sampling/sampling_logp_difference/max': 1.7155027389526367, 'sampling/importance_sampling_ratio/min': 0.1798732727766037, 'sampling/importance_sampling_ratio/mean': 1.0239506959915161, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8268949091434479, 'clip_ratio/low_mean': 0.09817329840734601, 'clip_ratio/low_min': 0.09817329840734601, 'clip_ratio/high_mean': 0.03622641507536173, 'clip_ratio/high_max': 0.03622641507536173, 'clip_ratio/region_mean': 0.13439971348270774, 'reward_total_mean': 0.15713383257389069, 'reward_meter_mean': 0.24479930102825165, 'reward_meter_std': 0.3385663628578186, '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_lexical_plausibility_mean': 0.9592169523239136, 'reward_lexical_plausibility_std': 0.10458080470561981, 'reward_repeat_penalty_mean': 0.8317855000495911, 'reward_repeat_penalty_std': 0.17168274521827698, 'reward_near_duplicate_penalty_mean': 0.9257292747497559, 'reward_near_duplicate_penalty_std': 0.04465325176715851, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9487742185592651, 'reward_distinct_2_std': 0.07594232261180878, 'reward_judge_quality_mean': 0.3712499737739563, 'reward_judge_quality_std': 0.25670647621154785, 'reward_total_composite_mean': 0.15713383257389069, 'reward_total_composite_std': 0.29908961057662964, 'epoch': 0.03}

 23%|██▎       | 767/3300 [3:03:46<9:46:58, 13.90s/it]INFO 04-12 13:49:59 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 768/3300 [3:03:55<8:44:20, 12.43s/it]2026-04-12 13:50:07,940 | INFO | train_grpo_train | metrics_logged mode=train step=768

                                                      
{'loss': 0.0237, 'grad_norm': 8.035225868225098, 'learning_rate': 7.675757575757577e-06, 'num_tokens': 1563614.0, 'completions/mean_length': 99.75, 'completions/min_length': 92.0, 'completions/max_length': 108.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 99.75, 'completions/min_terminated_length': 92.0, 'completions/max_terminated_length': 108.0, 'rewards/meter/mean': 0.7384893894195557, 'rewards/meter/std': 0.3352949023246765, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.42124998569488525, 'rewards/judge_quality/std': 0.22177450358867645, 'rewards/repeat_penalty/mean': 0.9049655199050903, 'rewards/repeat_penalty/std': 0.0951816514134407, 'rewards/near_duplicate_penalty/mean': 0.9490857124328613, 'rewards/near_duplicate_penalty/std': 0.024631649255752563, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9522029757499695, 'rewards/distinct_2/std': 0.08318986743688583, 'rewards/total_composite/mean': 0.3147892653942108, 'rewards/total_composite/std': 0.25307226181030273, 'reward': 0.3147892653942108, 'reward_std': 0.25307226181030273, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12582404911518097, 'sampling/sampling_logp_difference/max': 1.425328016281128, 'sampling/importance_sampling_ratio/min': 0.2404295951128006, 'sampling/importance_sampling_ratio/mean': 1.010919451713562, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9234515354037285, 'clip_ratio/low_mean': 0.055440484546124935, 'clip_ratio/low_min': 0.055440484546124935, 'clip_ratio/high_mean': 0.0771956229582429, 'clip_ratio/high_max': 0.0771956229582429, 'clip_ratio/region_mean': 0.13263610750436783, 'reward_total_mean': 0.3147892653942108, 'reward_meter_mean': 0.7384893894195557, 'reward_meter_std': 0.3352949023246765, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9049655199050903, 'reward_repeat_penalty_std': 0.0951816514134407, 'reward_near_duplicate_penalty_mean': 0.9490857124328613, 'reward_near_duplicate_penalty_std': 0.024631649255752563, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9522029757499695, 'reward_distinct_2_std': 0.08318986743688583, 'reward_judge_quality_mean': 0.42124998569488525, 'reward_judge_quality_std': 0.22177450358867645, 'reward_total_composite_mean': 0.3147892653942108, 'reward_total_composite_std': 0.25307226181030273, 'epoch': 0.03}

 23%|██▎       | 768/3300 [3:03:55<8:44:20, 12.43s/it]INFO 04-12 13:50:08 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 769/3300 [3:04:10<9:11:44, 13.08s/it]2026-04-12 13:50:22,546 | INFO | train_grpo_train | metrics_logged mode=train step=769

                                                      
{'loss': -0.0771, 'grad_norm': 4.437525749206543, 'learning_rate': 7.672727272727273e-06, 'num_tokens': 1566104.0, 'completions/mean_length': 190.25, 'completions/min_length': 125.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 144.2857208251953, 'completions/min_terminated_length': 125.0, 'completions/max_terminated_length': 166.0, 'rewards/meter/mean': 0.4939690828323364, 'rewards/meter/std': 0.37506309151649475, 'rewards/count_adherence/mean': 0.8500000238418579, 'rewards/count_adherence/std': 0.3505098223686218, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9726550579071045, 'rewards/lexical_plausibility/std': 0.06802557408809662, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.12464234977960587, 'rewards/repeat_penalty/mean': 0.806496798992157, 'rewards/repeat_penalty/std': 0.20518775284290314, 'rewards/near_duplicate_penalty/mean': 0.9333205223083496, 'rewards/near_duplicate_penalty/std': 0.0811675414443016, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8586480617523193, 'rewards/distinct_2/std': 0.157771036028862, 'rewards/total_composite/mean': 0.1271776258945465, 'rewards/total_composite/std': 0.10381483286619186, 'reward': 0.1271776258945465, 'reward_std': 0.10381483286619186, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1216985359787941, 'sampling/sampling_logp_difference/max': 1.6888399124145508, 'sampling/importance_sampling_ratio/min': 0.18473370373249054, 'sampling/importance_sampling_ratio/mean': 1.0117864608764648, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7788887470960617, 'clip_ratio/low_mean': 0.06171979568898678, 'clip_ratio/low_min': 0.06171979568898678, 'clip_ratio/high_mean': 0.03720343206077814, 'clip_ratio/high_max': 0.03720343206077814, 'clip_ratio/region_mean': 0.09892322774976492, 'reward_total_mean': 0.1271776258945465, 'reward_meter_mean': 0.4939690828323364, 'reward_meter_std': 0.37506309151649475, 'reward_count_adherence_mean': 0.8500000238418579, 'reward_count_adherence_std': 0.3505098223686218, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9726550579071045, 'reward_lexical_plausibility_std': 0.06802557408809662, 'reward_repeat_penalty_mean': 0.806496798992157, 'reward_repeat_penalty_std': 0.20518775284290314, 'reward_near_duplicate_penalty_mean': 0.9333205223083496, 'reward_near_duplicate_penalty_std': 0.0811675414443016, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8586480617523193, 'reward_distinct_2_std': 0.157771036028862, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.12464234977960587, 'reward_total_composite_mean': 0.1271776258945465, 'reward_total_composite_std': 0.10381483286619186, 'epoch': 0.03}

 23%|██▎       | 769/3300 [3:04:10<9:11:44, 13.08s/it]INFO 04-12 13:50:22 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 770/3300 [3:04:24<9:26:39, 13.44s/it]2026-04-12 13:50:36,824 | INFO | train_grpo_train | metrics_logged mode=train step=770

                                                      
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 7.66969696969697e-06, 'num_tokens': 1568096.0, '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': 0.8071956634521484, 'rewards/meter/std': 0.2763350307941437, 'rewards/count_adherence/mean': 0.8026316165924072, 'rewards/count_adherence/std': 0.06745998561382294, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.13125000894069672, 'rewards/judge_quality/std': 0.0530330128967762, 'rewards/repeat_penalty/mean': 0.14564862847328186, 'rewards/repeat_penalty/std': 0.22092342376708984, 'rewards/near_duplicate_penalty/mean': 0.721878707408905, 'rewards/near_duplicate_penalty/std': 0.12384537607431412, 'rewards/opening_diversity/mean': 0.9097259044647217, 'rewards/opening_diversity/std': 0.159354567527771, 'rewards/distinct_2/mean': 0.18396621942520142, 'rewards/distinct_2/std': 0.2518809139728546, 'rewards/total_composite/mean': 0.08450785279273987, 'rewards/total_composite/std': 0.04974975064396858, 'reward': 0.08450785279273987, 'reward_std': 0.049749743193387985, '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': 0.08450785279273987, 'reward_meter_mean': 0.8071956634521484, 'reward_meter_std': 0.2763350307941437, 'reward_count_adherence_mean': 0.8026316165924072, 'reward_count_adherence_std': 0.06745998561382294, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.14564862847328186, 'reward_repeat_penalty_std': 0.22092342376708984, 'reward_near_duplicate_penalty_mean': 0.721878707408905, 'reward_near_duplicate_penalty_std': 0.12384537607431412, 'reward_opening_diversity_mean': 0.9097259044647217, 'reward_opening_diversity_std': 0.159354567527771, 'reward_distinct_2_mean': 0.18396621942520142, 'reward_distinct_2_std': 0.2518809139728546, 'reward_judge_quality_mean': 0.13125000894069672, 'reward_judge_quality_std': 0.0530330128967762, 'reward_total_composite_mean': 0.08450785279273987, 'reward_total_composite_std': 0.04974975064396858, 'epoch': 0.03}

 23%|██▎       | 770/3300 [3:04:24<9:26:39, 13.44s/it]INFO 04-12 13:50:37 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 771/3300 [3:04:39<9:40:56, 13.78s/it]2026-04-12 13:50:51,411 | INFO | train_grpo_train | metrics_logged mode=train step=771

                                                      
{'loss': -0.2015, 'grad_norm': 2.096440553665161, 'learning_rate': 7.666666666666667e-06, 'num_tokens': 1570330.0, 'completions/mean_length': 335.25, 'completions/min_length': 156.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 158.5, 'completions/min_terminated_length': 156.0, 'completions/max_terminated_length': 161.0, 'rewards/meter/mean': 0.7819421887397766, 'rewards/meter/std': 0.3578019142150879, 'rewards/count_adherence/mean': 0.7749999761581421, 'rewards/count_adherence/std': 0.328416109085083, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/lexical_plausibility/mean': 0.8504629731178284, 'rewards/lexical_plausibility/std': 0.18675591051578522, 'rewards/judge_quality/mean': 0.20000000298023224, 'rewards/judge_quality/std': 0.16035674512386322, 'rewards/repeat_penalty/mean': 0.6678902506828308, 'rewards/repeat_penalty/std': 0.32416048645973206, 'rewards/near_duplicate_penalty/mean': 0.8827751874923706, 'rewards/near_duplicate_penalty/std': 0.11507882177829742, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.15221577882766724, 'rewards/distinct_2/mean': 0.792546272277832, 'rewards/distinct_2/std': 0.2205480933189392, 'rewards/total_composite/mean': 0.14750972390174866, 'rewards/total_composite/std': 0.17079360783100128, 'reward': 0.14750972390174866, 'reward_std': 0.17079360783100128, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10303334891796112, 'sampling/sampling_logp_difference/max': 1.0576629638671875, 'sampling/importance_sampling_ratio/min': 0.34726640582084656, 'sampling/importance_sampling_ratio/mean': 1.025027871131897, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4936429560184479, 'clip_ratio/low_mean': 0.009316770359873772, 'clip_ratio/low_min': 0.009316770359873772, 'clip_ratio/high_mean': 0.03644225653260946, 'clip_ratio/high_max': 0.03644225653260946, 'clip_ratio/region_mean': 0.045759026892483234, 'reward_total_mean': 0.14750972390174866, 'reward_meter_mean': 0.7819421887397766, 'reward_meter_std': 0.3578019142150879, 'reward_count_adherence_mean': 0.7749999761581421, 'reward_count_adherence_std': 0.328416109085083, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_lexical_plausibility_mean': 0.8504629731178284, 'reward_lexical_plausibility_std': 0.18675591051578522, 'reward_repeat_penalty_mean': 0.6678902506828308, 'reward_repeat_penalty_std': 0.32416048645973206, 'reward_near_duplicate_penalty_mean': 0.8827751874923706, 'reward_near_duplicate_penalty_std': 0.11507882177829742, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.15221577882766724, 'reward_distinct_2_mean': 0.792546272277832, 'reward_distinct_2_std': 0.2205480933189392, 'reward_judge_quality_mean': 0.20000000298023224, 'reward_judge_quality_std': 0.16035674512386322, 'reward_total_composite_mean': 0.14750972390174866, 'reward_total_composite_std': 0.17079360783100128, 'epoch': 0.03}

 23%|██▎       | 771/3300 [3:04:39<9:40:56, 13.78s/it]INFO 04-12 13:50:51 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 772/3300 [3:04:49<8:51:59, 12.63s/it]2026-04-12 13:51:01,342 | INFO | train_grpo_train | metrics_logged mode=train step=772

                                                      
{'loss': -0.0472, 'grad_norm': 7.703415870666504, 'learning_rate': 7.663636363636364e-06, 'num_tokens': 1572507.0, 'completions/mean_length': 108.125, 'completions/min_length': 70.0, 'completions/max_length': 119.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 108.125, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 119.0, 'rewards/meter/mean': 0.42608457803726196, 'rewards/meter/std': 0.3771738111972809, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.0353553369641304, 'rewards/repeat_penalty/mean': 0.6464917063713074, 'rewards/repeat_penalty/std': 0.16450101137161255, 'rewards/near_duplicate_penalty/mean': 0.9018651247024536, 'rewards/near_duplicate_penalty/std': 0.037164103239774704, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.7270519733428955, 'rewards/distinct_2/std': 0.14556081593036652, 'rewards/total_composite/mean': 0.1489199697971344, 'rewards/total_composite/std': 0.13219745457172394, 'reward': 0.1489199697971344, 'reward_std': 0.13219745457172394, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10423891246318817, 'sampling/sampling_logp_difference/max': 1.7511940002441406, 'sampling/importance_sampling_ratio/min': 0.17356657981872559, 'sampling/importance_sampling_ratio/mean': 1.0041307210922241, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7611419856548309, 'clip_ratio/low_mean': 0.051392852794378996, 'clip_ratio/low_min': 0.051392852794378996, 'clip_ratio/high_mean': 0.05071981530636549, 'clip_ratio/high_max': 0.05071981530636549, 'clip_ratio/region_mean': 0.10211266810074449, 'reward_total_mean': 0.1489199697971344, 'reward_meter_mean': 0.42608457803726196, 'reward_meter_std': 0.3771738111972809, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6464917063713074, 'reward_repeat_penalty_std': 0.16450101137161255, 'reward_near_duplicate_penalty_mean': 0.9018651247024536, 'reward_near_duplicate_penalty_std': 0.037164103239774704, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.7270519733428955, 'reward_distinct_2_std': 0.14556081593036652, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.0353553369641304, 'reward_total_composite_mean': 0.1489199697971344, 'reward_total_composite_std': 0.13219745457172394, 'epoch': 0.03}

 23%|██▎       | 772/3300 [3:04:49<8:51:59, 12.63s/it]INFO 04-12 13:51:01 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 773/3300 [3:05:03<9:16:59, 13.22s/it]2026-04-12 13:51:15,963 | INFO | train_grpo_train | metrics_logged mode=train step=773

                                                      
{'loss': -0.2456, 'grad_norm': 1.5966325998306274, 'learning_rate': 7.660606060606062e-06, 'num_tokens': 1574963.0, 'completions/mean_length': 323.0, 'completions/min_length': 179.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 209.60000610351562, 'completions/min_terminated_length': 179.0, 'completions/max_terminated_length': 228.0, 'rewards/meter/mean': 0.7656537294387817, 'rewards/meter/std': 0.3421063721179962, 'rewards/count_adherence/mean': 0.7678571343421936, 'rewards/count_adherence/std': 0.3322991728782654, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.8451470732688904, 'rewards/lexical_plausibility/std': 0.2173515111207962, 'rewards/judge_quality/mean': 0.11249999701976776, 'rewards/judge_quality/std': 0.051754921674728394, 'rewards/repeat_penalty/mean': 0.4553380012512207, 'rewards/repeat_penalty/std': 0.4441148638725281, 'rewards/near_duplicate_penalty/mean': 0.8036165237426758, 'rewards/near_duplicate_penalty/std': 0.17423562705516815, 'rewards/opening_diversity/mean': 0.9642857313156128, 'rewards/opening_diversity/std': 0.06613000482320786, 'rewards/distinct_2/mean': 0.5186556577682495, 'rewards/distinct_2/std': 0.4284680485725403, 'rewards/total_composite/mean': 0.08388496935367584, 'rewards/total_composite/std': 0.07035890966653824, 'reward': 0.08388496935367584, 'reward_std': 0.07035890966653824, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0626145452260971, 'sampling/sampling_logp_difference/max': 1.1764965057373047, 'sampling/importance_sampling_ratio/min': 0.3083571791648865, 'sampling/importance_sampling_ratio/mean': 1.010095477104187, 'sampling/importance_sampling_ratio/max': 1.9337958097457886, 'entropy': 0.3157663568854332, 'clip_ratio/low_mean': 0.008223684504628181, 'clip_ratio/low_min': 0.008223684504628181, 'clip_ratio/high_mean': 0.03121564514003694, 'clip_ratio/high_max': 0.03121564514003694, 'clip_ratio/region_mean': 0.03943932964466512, 'reward_total_mean': 0.08388496935367584, 'reward_meter_mean': 0.7656537294387817, 'reward_meter_std': 0.3421063721179962, 'reward_count_adherence_mean': 0.7678571343421936, 'reward_count_adherence_std': 0.3322991728782654, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.8451470732688904, 'reward_lexical_plausibility_std': 0.2173515111207962, 'reward_repeat_penalty_mean': 0.4553380012512207, 'reward_repeat_penalty_std': 0.4441148638725281, 'reward_near_duplicate_penalty_mean': 0.8036165237426758, 'reward_near_duplicate_penalty_std': 0.17423562705516815, 'reward_opening_diversity_mean': 0.9642857313156128, 'reward_opening_diversity_std': 0.06613000482320786, 'reward_distinct_2_mean': 0.5186556577682495, 'reward_distinct_2_std': 0.4284680485725403, 'reward_judge_quality_mean': 0.11249999701976776, 'reward_judge_quality_std': 0.051754921674728394, 'reward_total_composite_mean': 0.08388496935367584, 'reward_total_composite_std': 0.07035890966653824, 'epoch': 0.03}

 23%|██▎       | 773/3300 [3:05:03<9:16:59, 13.22s/it]INFO 04-12 13:51:16 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 774/3300 [3:05:17<9:22:46, 13.37s/it]2026-04-12 13:51:29,661 | INFO | train_grpo_train | metrics_logged mode=train step=774

                                                      
{'loss': 0.6796, 'grad_norm': 8.25854206085205, 'learning_rate': 7.657575757575757e-06, 'num_tokens': 1577102.0, 'completions/mean_length': 104.375, 'completions/min_length': 48.0, 'completions/max_length': 456.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 104.375, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 456.0, 'rewards/meter/mean': 0.6154877543449402, 'rewards/meter/std': 0.3558361530303955, '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/lexical_plausibility/mean': 0.9594399929046631, 'rewards/lexical_plausibility/std': 0.114720918238163, 'rewards/judge_quality/mean': 0.35874998569488525, 'rewards/judge_quality/std': 0.27105283737182617, 'rewards/repeat_penalty/mean': 0.891266405582428, 'rewards/repeat_penalty/std': 0.13831356167793274, 'rewards/near_duplicate_penalty/mean': 0.951568603515625, 'rewards/near_duplicate_penalty/std': 0.03288222849369049, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9604700803756714, 'rewards/distinct_2/std': 0.059027865529060364, 'rewards/total_composite/mean': 0.21076039969921112, 'rewards/total_composite/std': 0.15587833523750305, 'reward': 0.21076039969921112, 'reward_std': 0.15587833523750305, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14122696220874786, 'sampling/sampling_logp_difference/max': 1.2438321113586426, 'sampling/importance_sampling_ratio/min': 0.28827738761901855, 'sampling/importance_sampling_ratio/mean': 1.0190385580062866, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.172049954533577, 'clip_ratio/low_mean': 0.0631115126889199, 'clip_ratio/low_min': 0.0631115126889199, 'clip_ratio/high_mean': 0.053981839679181576, 'clip_ratio/high_max': 0.053981839679181576, 'clip_ratio/region_mean': 0.11709335236810148, 'reward_total_mean': 0.21076039969921112, 'reward_meter_mean': 0.6154877543449402, 'reward_meter_std': 0.3558361530303955, '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_lexical_plausibility_mean': 0.9594399929046631, 'reward_lexical_plausibility_std': 0.114720918238163, 'reward_repeat_penalty_mean': 0.891266405582428, 'reward_repeat_penalty_std': 0.13831356167793274, 'reward_near_duplicate_penalty_mean': 0.951568603515625, 'reward_near_duplicate_penalty_std': 0.03288222849369049, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9604700803756714, 'reward_distinct_2_std': 0.059027865529060364, 'reward_judge_quality_mean': 0.35874998569488525, 'reward_judge_quality_std': 0.27105283737182617, 'reward_total_composite_mean': 0.21076039969921112, 'reward_total_composite_std': 0.15587833523750305, 'epoch': 0.03}

 23%|██▎       | 774/3300 [3:05:17<9:22:46, 13.37s/it]INFO 04-12 13:51:29 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 775/3300 [3:05:28<8:46:16, 12.51s/it]2026-04-12 13:51:40,156 | INFO | train_grpo_train | metrics_logged mode=train step=775

                                                      
{'loss': 0.0225, 'grad_norm': 7.958051681518555, 'learning_rate': 7.654545454545456e-06, 'num_tokens': 1579360.0, 'completions/mean_length': 116.25, 'completions/min_length': 93.0, 'completions/max_length': 131.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 116.25, 'completions/min_terminated_length': 93.0, 'completions/max_terminated_length': 131.0, 'rewards/meter/mean': 0.9873641133308411, 'rewards/meter/std': 0.006833996623754501, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.08345229178667068, 'rewards/repeat_penalty/mean': 0.5505112409591675, 'rewards/repeat_penalty/std': 0.24144576489925385, 'rewards/near_duplicate_penalty/mean': 0.8619503974914551, 'rewards/near_duplicate_penalty/std': 0.08173158764839172, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.6537202596664429, 'rewards/distinct_2/std': 0.21680009365081787, 'rewards/total_composite/mean': 0.32486623525619507, 'rewards/total_composite/std': 0.08514490723609924, 'reward': 0.32486623525619507, 'reward_std': 0.08514490723609924, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1114606112241745, 'sampling/sampling_logp_difference/max': 2.142119884490967, 'sampling/importance_sampling_ratio/min': 0.11740569770336151, 'sampling/importance_sampling_ratio/mean': 1.004138469696045, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5601413436233997, 'clip_ratio/low_mean': 0.01877616299316287, 'clip_ratio/low_min': 0.01877616299316287, 'clip_ratio/high_mean': 0.07950412016361952, 'clip_ratio/high_max': 0.07950412016361952, 'clip_ratio/region_mean': 0.09828028315678239, 'reward_total_mean': 0.32486623525619507, 'reward_meter_mean': 0.9873641133308411, 'reward_meter_std': 0.006833996623754501, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.5505112409591675, 'reward_repeat_penalty_std': 0.24144576489925385, 'reward_near_duplicate_penalty_mean': 0.8619503974914551, 'reward_near_duplicate_penalty_std': 0.08173158764839172, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.6537202596664429, 'reward_distinct_2_std': 0.21680009365081787, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.08345229178667068, 'reward_total_composite_mean': 0.32486623525619507, 'reward_total_composite_std': 0.08514490723609924, 'epoch': 0.03}

 23%|██▎       | 775/3300 [3:05:28<8:46:16, 12.51s/it]INFO 04-12 13:51:40 [block_pool.py:378] Successfully reset prefix cache

 24%|██▎       | 776/3300 [3:05:42<9:11:13, 13.10s/it]2026-04-12 13:51:54,655 | INFO | train_grpo_train | metrics_logged mode=train step=776

                                                      
{'loss': -0.0931, 'grad_norm': 3.8162519931793213, 'learning_rate': 7.651515151515152e-06, 'num_tokens': 1580974.0, 'completions/mean_length': 110.75, 'completions/min_length': 46.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 53.42857360839844, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.6486303210258484, 'rewards/meter/std': 0.33514273166656494, '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/lexical_plausibility/mean': 0.9562513828277588, 'rewards/lexical_plausibility/std': 0.12373967468738556, 'rewards/judge_quality/mean': 0.39625000953674316, 'rewards/judge_quality/std': 0.2585087716579437, 'rewards/repeat_penalty/mean': 0.8737409114837646, 'rewards/repeat_penalty/std': 0.07075395435094833, 'rewards/near_duplicate_penalty/mean': 0.891111433506012, 'rewards/near_duplicate_penalty/std': 0.06787948310375214, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9805935621261597, 'rewards/distinct_2/std': 0.03112947940826416, 'rewards/total_composite/mean': 0.2499900758266449, 'rewards/total_composite/std': 0.14878235757350922, 'reward': 0.2499900758266449, 'reward_std': 0.14878235757350922, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12070675194263458, 'sampling/sampling_logp_difference/max': 1.8354015350341797, 'sampling/importance_sampling_ratio/min': 0.15954943001270294, 'sampling/importance_sampling_ratio/mean': 1.0159733295440674, 'sampling/importance_sampling_ratio/max': 1.824219822883606, 'entropy': 0.6956275403499603, 'clip_ratio/low_mean': 0.01854395680129528, 'clip_ratio/low_min': 0.01854395680129528, 'clip_ratio/high_mean': 0.06450344668701291, 'clip_ratio/high_max': 0.06450344668701291, 'clip_ratio/region_mean': 0.08304740348830819, 'reward_total_mean': 0.2499900758266449, 'reward_meter_mean': 0.6486303210258484, 'reward_meter_std': 0.33514273166656494, '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_lexical_plausibility_mean': 0.9562513828277588, 'reward_lexical_plausibility_std': 0.12373967468738556, 'reward_repeat_penalty_mean': 0.8737409114837646, 'reward_repeat_penalty_std': 0.07075395435094833, 'reward_near_duplicate_penalty_mean': 0.891111433506012, 'reward_near_duplicate_penalty_std': 0.06787948310375214, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9805935621261597, 'reward_distinct_2_std': 0.03112947940826416, 'reward_judge_quality_mean': 0.39625000953674316, 'reward_judge_quality_std': 0.2585087716579437, 'reward_total_composite_mean': 0.2499900758266449, 'reward_total_composite_std': 0.14878235757350922, 'epoch': 0.03}

 24%|██▎       | 776/3300 [3:05:42<9:11:13, 13.10s/it]INFO 04-12 13:51:54 [block_pool.py:378] Successfully reset prefix cache

 24%|██▎       | 777/3300 [3:05:50<8:06:10, 11.56s/it]2026-04-12 13:52:02,615 | INFO | train_grpo_train | metrics_logged mode=train step=777

                                                      
{'loss': 0.0398, 'grad_norm': 16.43800926208496, 'learning_rate': 7.648484848484849e-06, 'num_tokens': 1582393.0, 'completions/mean_length': 24.375, 'completions/min_length': 21.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.375, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.7629504203796387, 'rewards/meter/std': 0.40981799364089966, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.41749998927116394, 'rewards/judge_quality/std': 0.32805708050727844, 'rewards/repeat_penalty/mean': 0.7214998006820679, 'rewards/repeat_penalty/std': 0.043466612696647644, 'rewards/near_duplicate_penalty/mean': 0.961999773979187, 'rewards/near_duplicate_penalty/std': 0.057955484837293625, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2467324286699295, 'rewards/total_composite/std': 0.3028090000152588, 'reward': 0.2467324286699295, 'reward_std': 0.3028090000152588, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16168943047523499, 'sampling/sampling_logp_difference/max': 1.7030162811279297, 'sampling/importance_sampling_ratio/min': 0.18213333189487457, 'sampling/importance_sampling_ratio/mean': 1.0011332035064697, 'sampling/importance_sampling_ratio/max': 1.947780966758728, 'entropy': 1.216828003525734, 'clip_ratio/low_mean': 0.09256851579993963, 'clip_ratio/low_min': 0.09256851579993963, 'clip_ratio/high_mean': 0.06101190485060215, 'clip_ratio/high_max': 0.06101190485060215, 'clip_ratio/region_mean': 0.15358042065054178, 'reward_total_mean': 0.2467324286699295, 'reward_meter_mean': 0.7629504203796387, 'reward_meter_std': 0.40981799364089966, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7214998006820679, 'reward_repeat_penalty_std': 0.043466612696647644, 'reward_near_duplicate_penalty_mean': 0.961999773979187, 'reward_near_duplicate_penalty_std': 0.057955484837293625, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.41749998927116394, 'reward_judge_quality_std': 0.32805708050727844, 'reward_total_composite_mean': 0.2467324286699295, 'reward_total_composite_std': 0.3028090000152588, 'epoch': 0.03}

 24%|██▎       | 777/3300 [3:05:50<8:06:10, 11.56s/it]INFO 04-12 13:52:02 [block_pool.py:378] Successfully reset prefix cache

 24%|██▎       | 778/3300 [3:05:59<7:31:23, 10.74s/it]2026-04-12 13:52:11,437 | INFO | train_grpo_train | metrics_logged mode=train step=778

                                                      
{'loss': 0.0589, 'grad_norm': 9.360218048095703, 'learning_rate': 7.645454545454546e-06, 'num_tokens': 1584393.0, 'completions/mean_length': 69.0, 'completions/min_length': 64.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.9810389280319214, 'rewards/meter/std': 0.011988348327577114, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5512499809265137, 'rewards/judge_quality/std': 0.31638303399086, 'rewards/repeat_penalty/mean': 0.7110658884048462, 'rewards/repeat_penalty/std': 0.3008081614971161, 'rewards/near_duplicate_penalty/mean': 0.8966309428215027, 'rewards/near_duplicate_penalty/std': 0.08771822601556778, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.8544689416885376, 'rewards/distinct_2/std': 0.1578684151172638, 'rewards/total_composite/mean': 0.519660234451294, 'rewards/total_composite/std': 0.3348424434661865, 'reward': 0.519660234451294, 'reward_std': 0.33484241366386414, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1279457062482834, 'sampling/sampling_logp_difference/max': 1.3275394439697266, 'sampling/importance_sampling_ratio/min': 0.26512885093688965, 'sampling/importance_sampling_ratio/mean': 1.0034544467926025, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.880782924592495, 'clip_ratio/low_mean': 0.06230261968448758, 'clip_ratio/low_min': 0.06230261968448758, 'clip_ratio/high_mean': 0.07294372469186783, 'clip_ratio/high_max': 0.07294372469186783, 'clip_ratio/region_mean': 0.1352463443763554, 'reward_total_mean': 0.519660234451294, 'reward_meter_mean': 0.9810389280319214, 'reward_meter_std': 0.011988348327577114, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7110658884048462, 'reward_repeat_penalty_std': 0.3008081614971161, 'reward_near_duplicate_penalty_mean': 0.8966309428215027, 'reward_near_duplicate_penalty_std': 0.08771822601556778, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.8544689416885376, 'reward_distinct_2_std': 0.1578684151172638, 'reward_judge_quality_mean': 0.5512499809265137, 'reward_judge_quality_std': 0.31638303399086, 'reward_total_composite_mean': 0.519660234451294, 'reward_total_composite_std': 0.3348424434661865, 'epoch': 0.03}

 24%|██▎       | 778/3300 [3:05:59<7:31:23, 10.74s/it]INFO 04-12 13:52:11 [block_pool.py:378] Successfully reset prefix cache

 24%|██▎       | 779/3300 [3:06:12<8:05:22, 11.55s/it]2026-04-12 13:52:24,887 | INFO | train_grpo_train | metrics_logged mode=train step=779

                                                      
{'loss': -0.0641, 'grad_norm': 4.622972011566162, 'learning_rate': 7.642424242424244e-06, 'num_tokens': 1585794.0, 'completions/mean_length': 102.125, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 43.57143020629883, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.6433348655700684, 'rewards/meter/std': 0.35852622985839844, '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/lexical_plausibility/mean': 0.9826866388320923, 'rewards/lexical_plausibility/std': 0.04896962642669678, 'rewards/judge_quality/mean': 0.5962499976158142, 'rewards/judge_quality/std': 0.32945573329925537, 'rewards/repeat_penalty/mean': 0.9430086016654968, 'rewards/repeat_penalty/std': 0.08734755218029022, 'rewards/near_duplicate_penalty/mean': 0.9601523876190186, 'rewards/near_duplicate_penalty/std': 0.05086527392268181, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9807692170143127, 'rewards/distinct_2/std': 0.054392825812101364, 'rewards/total_composite/mean': 0.392836332321167, 'rewards/total_composite/std': 0.2896917760372162, 'reward': 0.392836332321167, 'reward_std': 0.2896917760372162, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13461820781230927, 'sampling/sampling_logp_difference/max': 1.3629393577575684, 'sampling/importance_sampling_ratio/min': 0.25590747594833374, 'sampling/importance_sampling_ratio/mean': 1.0438573360443115, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6360959336161613, 'clip_ratio/low_mean': 0.03655368648469448, 'clip_ratio/low_min': 0.03655368648469448, 'clip_ratio/high_mean': 0.061165111139416695, 'clip_ratio/high_max': 0.061165111139416695, 'clip_ratio/region_mean': 0.09771879762411118, 'reward_total_mean': 0.392836332321167, 'reward_meter_mean': 0.6433348655700684, 'reward_meter_std': 0.35852622985839844, '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_lexical_plausibility_mean': 0.9826866388320923, 'reward_lexical_plausibility_std': 0.04896962642669678, 'reward_repeat_penalty_mean': 0.9430086016654968, 'reward_repeat_penalty_std': 0.08734755218029022, 'reward_near_duplicate_penalty_mean': 0.9601523876190186, 'reward_near_duplicate_penalty_std': 0.05086527392268181, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9807692170143127, 'reward_distinct_2_std': 0.054392825812101364, 'reward_judge_quality_mean': 0.5962499976158142, 'reward_judge_quality_std': 0.32945573329925537, 'reward_total_composite_mean': 0.392836332321167, 'reward_total_composite_std': 0.2896917760372162, 'epoch': 0.03}

 24%|██▎       | 779/3300 [3:06:12<8:05:22, 11.55s/it]INFO 04-12 13:52:25 [block_pool.py:378] Successfully reset prefix cache

 24%|██▎       | 780/3300 [3:06:21<7:35:26, 10.84s/it]2026-04-12 13:52:34,076 | INFO | train_grpo_train | metrics_logged mode=train step=780

                                                      
{'loss': 0.0711, 'grad_norm': 9.191865921020508, 'learning_rate': 7.639393939393939e-06, 'num_tokens': 1587987.0, 'completions/mean_length': 87.125, 'completions/min_length': 69.0, 'completions/max_length': 111.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 87.125, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 111.0, 'rewards/meter/mean': 0.4510396122932434, 'rewards/meter/std': 0.3611578941345215, 'rewards/count_adherence/mean': 0.84375, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.41249996423721313, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.8547777533531189, 'rewards/repeat_penalty/std': 0.10875024646520615, 'rewards/near_duplicate_penalty/mean': 0.9258180856704712, 'rewards/near_duplicate_penalty/std': 0.06282761693000793, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9203787446022034, 'rewards/distinct_2/std': 0.06352870911359787, 'rewards/total_composite/mean': 0.15372705459594727, 'rewards/total_composite/std': 0.12295910716056824, 'reward': 0.15372705459594727, 'reward_std': 0.12295910716056824, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10808668285608292, 'sampling/sampling_logp_difference/max': 1.3506691455841064, 'sampling/importance_sampling_ratio/min': 0.2590668499469757, 'sampling/importance_sampling_ratio/mean': 1.0184504985809326, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7183883041143417, 'clip_ratio/low_mean': 0.05325394216924906, 'clip_ratio/low_min': 0.05325394216924906, 'clip_ratio/high_mean': 0.06475755199790001, 'clip_ratio/high_max': 0.06475755199790001, 'clip_ratio/region_mean': 0.11801149416714907, 'reward_total_mean': 0.15372705459594727, 'reward_meter_mean': 0.4510396122932434, 'reward_meter_std': 0.3611578941345215, 'reward_count_adherence_mean': 0.84375, 'reward_count_adherence_std': 0.12938730418682098, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8547777533531189, 'reward_repeat_penalty_std': 0.10875024646520615, 'reward_near_duplicate_penalty_mean': 0.9258180856704712, 'reward_near_duplicate_penalty_std': 0.06282761693000793, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9203787446022034, 'reward_distinct_2_std': 0.06352870911359787, 'reward_judge_quality_mean': 0.41249996423721313, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.15372705459594727, 'reward_total_composite_std': 0.12295910716056824, 'epoch': 0.03}

 24%|██▎       | 780/3300 [3:06:22<7:35:26, 10.84s/it]INFO 04-12 13:52:34 [block_pool.py:378] Successfully reset prefix cache

 24%|██▎       | 781/3300 [3:06:30<7:00:38, 10.02s/it]2026-04-12 13:52:42,171 | INFO | train_grpo_train | metrics_logged mode=train step=781

                                                      
{'loss': 0.061, 'grad_norm': 16.4144344329834, 'learning_rate': 7.636363636363638e-06, 'num_tokens': 1589590.0, 'completions/mean_length': 38.375, 'completions/min_length': 35.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.375, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.4185240566730499, 'rewards/meter/std': 0.4068979322910309, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9657738208770752, 'rewards/lexical_plausibility/std': 0.06534077972173691, 'rewards/judge_quality/mean': 0.8524999618530273, 'rewards/judge_quality/std': 0.20331187546253204, 'rewards/repeat_penalty/mean': 0.9121558666229248, 'rewards/repeat_penalty/std': 0.0743054747581482, 'rewards/near_duplicate_penalty/mean': 0.935652494430542, 'rewards/near_duplicate_penalty/std': 0.05105455219745636, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9743589758872986, 'rewards/distinct_2/std': 0.047477953135967255, 'rewards/total_composite/mean': 0.38194817304611206, 'rewards/total_composite/std': 0.38644152879714966, 'reward': 0.38194817304611206, 'reward_std': 0.38644155859947205, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17882081866264343, 'sampling/sampling_logp_difference/max': 2.011857032775879, 'sampling/importance_sampling_ratio/min': 0.13374008238315582, 'sampling/importance_sampling_ratio/mean': 1.026328206062317, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8333993703126907, 'clip_ratio/low_mean': 0.06678113713860512, 'clip_ratio/low_min': 0.06678113713860512, 'clip_ratio/high_mean': 0.06350760627537966, 'clip_ratio/high_max': 0.06350760627537966, 'clip_ratio/region_mean': 0.13028874341398478, 'reward_total_mean': 0.38194817304611206, 'reward_meter_mean': 0.4185240566730499, 'reward_meter_std': 0.4068979322910309, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9657738208770752, 'reward_lexical_plausibility_std': 0.06534077972173691, 'reward_repeat_penalty_mean': 0.9121558666229248, 'reward_repeat_penalty_std': 0.0743054747581482, 'reward_near_duplicate_penalty_mean': 0.935652494430542, 'reward_near_duplicate_penalty_std': 0.05105455219745636, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9743589758872986, 'reward_distinct_2_std': 0.047477953135967255, 'reward_judge_quality_mean': 0.8524999618530273, 'reward_judge_quality_std': 0.20331187546253204, 'reward_total_composite_mean': 0.38194817304611206, 'reward_total_composite_std': 0.38644152879714966, 'epoch': 0.03}

 24%|██▎       | 781/3300 [3:06:30<7:00:38, 10.02s/it]INFO 04-12 13:52:42 [block_pool.py:378] Successfully reset prefix cache

 24%|██▎       | 782/3300 [3:06:38<6:45:09,  9.65s/it]2026-04-12 13:52:50,977 | INFO | train_grpo_train | metrics_logged mode=train step=782

                                                      
{'loss': 0.0893, 'grad_norm': 9.844320297241211, 'learning_rate': 7.633333333333334e-06, 'num_tokens': 1591514.0, 'completions/mean_length': 82.5, 'completions/min_length': 64.0, 'completions/max_length': 95.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 82.5, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 95.0, 'rewards/meter/mean': 0.785467803478241, 'rewards/meter/std': 0.3208889663219452, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4662500023841858, 'rewards/judge_quality/std': 0.17435084283351898, 'rewards/repeat_penalty/mean': 0.906518816947937, 'rewards/repeat_penalty/std': 0.13854116201400757, 'rewards/near_duplicate_penalty/mean': 0.9609131813049316, 'rewards/near_duplicate_penalty/std': 0.024943312630057335, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9670645594596863, 'rewards/distinct_2/std': 0.05618254095315933, 'rewards/total_composite/mean': 0.3451053500175476, 'rewards/total_composite/std': 0.15760640799999237, 'reward': 0.3451053500175476, 'reward_std': 0.15760640799999237, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11764537543058395, 'sampling/sampling_logp_difference/max': 1.815509557723999, 'sampling/importance_sampling_ratio/min': 0.2030559778213501, 'sampling/importance_sampling_ratio/mean': 1.0113693475723267, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7865246385335922, 'clip_ratio/low_mean': 0.04976032115519047, 'clip_ratio/low_min': 0.04976032115519047, 'clip_ratio/high_mean': 0.04963441099971533, 'clip_ratio/high_max': 0.04963441099971533, 'clip_ratio/region_mean': 0.0993947321549058, 'reward_total_mean': 0.3451053500175476, 'reward_meter_mean': 0.785467803478241, 'reward_meter_std': 0.3208889663219452, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.906518816947937, 'reward_repeat_penalty_std': 0.13854116201400757, 'reward_near_duplicate_penalty_mean': 0.9609131813049316, 'reward_near_duplicate_penalty_std': 0.024943312630057335, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9670645594596863, 'reward_distinct_2_std': 0.05618254095315933, 'reward_judge_quality_mean': 0.4662500023841858, 'reward_judge_quality_std': 0.17435084283351898, 'reward_total_composite_mean': 0.3451053500175476, 'reward_total_composite_std': 0.15760640799999237, 'epoch': 0.03}

 24%|██▎       | 782/3300 [3:06:38<6:45:09,  9.65s/it]INFO 04-12 13:52:51 [block_pool.py:378] Successfully reset prefix cache

 24%|██▎       | 783/3300 [3:06:48<6:39:53,  9.53s/it]2026-04-12 13:53:00,227 | INFO | train_grpo_train | metrics_logged mode=train step=783

                                                      
{'loss': 0.0563, 'grad_norm': 12.375184059143066, 'learning_rate': 7.630303030303031e-06, 'num_tokens': 1593360.0, 'completions/mean_length': 61.75, 'completions/min_length': 55.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.75, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.7023825645446777, 'rewards/meter/std': 0.2803703248500824, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.4962499737739563, 'rewards/judge_quality/std': 0.17476005852222443, 'rewards/repeat_penalty/mean': 0.8060828447341919, 'rewards/repeat_penalty/std': 0.16500318050384521, 'rewards/near_duplicate_penalty/mean': 0.9578722715377808, 'rewards/near_duplicate_penalty/std': 0.028928671032190323, 'rewards/opening_diversity/mean': 0.875, 'rewards/opening_diversity/std': 0.13363061845302582, 'rewards/distinct_2/mean': 0.9550883769989014, 'rewards/distinct_2/std': 0.0376155711710453, 'rewards/total_composite/mean': 0.35603511333465576, 'rewards/total_composite/std': 0.19789965450763702, 'reward': 0.35603511333465576, 'reward_std': 0.19789965450763702, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1304122656583786, 'sampling/sampling_logp_difference/max': 2.1529884338378906, 'sampling/importance_sampling_ratio/min': 0.1161365658044815, 'sampling/importance_sampling_ratio/mean': 1.0186083316802979, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9203171506524086, 'clip_ratio/low_mean': 0.04593327455222607, 'clip_ratio/low_min': 0.04593327455222607, 'clip_ratio/high_mean': 0.1020668912678957, 'clip_ratio/high_max': 0.1020668912678957, 'clip_ratio/region_mean': 0.14800016582012177, 'reward_total_mean': 0.35603511333465576, 'reward_meter_mean': 0.7023825645446777, 'reward_meter_std': 0.2803703248500824, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.8060828447341919, 'reward_repeat_penalty_std': 0.16500318050384521, 'reward_near_duplicate_penalty_mean': 0.9578722715377808, 'reward_near_duplicate_penalty_std': 0.028928671032190323, 'reward_opening_diversity_mean': 0.875, 'reward_opening_diversity_std': 0.13363061845302582, 'reward_distinct_2_mean': 0.9550883769989014, 'reward_distinct_2_std': 0.0376155711710453, 'reward_judge_quality_mean': 0.4962499737739563, 'reward_judge_quality_std': 0.17476005852222443, 'reward_total_composite_mean': 0.35603511333465576, 'reward_total_composite_std': 0.19789965450763702, 'epoch': 0.03}

 24%|██▎       | 783/3300 [3:06:48<6:39:53,  9.53s/it]INFO 04-12 13:53:00 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 784/3300 [3:06:57<6:31:18,  9.33s/it]2026-04-12 13:53:09,087 | INFO | train_grpo_train | metrics_logged mode=train step=784

                                                      
{'loss': 0.0805, 'grad_norm': 10.62993335723877, 'learning_rate': 7.627272727272727e-06, 'num_tokens': 1595054.0, 'completions/mean_length': 55.75, 'completions/min_length': 48.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.75, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.4122018814086914, 'rewards/meter/std': 0.45157957077026367, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.995192289352417, 'rewards/lexical_plausibility/std': 0.013598216697573662, 'rewards/judge_quality/mean': 0.47999998927116394, 'rewards/judge_quality/std': 0.28400203585624695, 'rewards/repeat_penalty/mean': 0.9630240797996521, 'rewards/repeat_penalty/std': 0.031217806041240692, 'rewards/near_duplicate_penalty/mean': 0.9630240797996521, 'rewards/near_duplicate_penalty/std': 0.031217806041240692, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.22415006160736084, 'rewards/total_composite/std': 0.30988043546676636, 'reward': 0.22415006160736084, 'reward_std': 0.30988043546676636, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13101045787334442, 'sampling/sampling_logp_difference/max': 2.3957738876342773, 'sampling/importance_sampling_ratio/min': 0.09110214561223984, 'sampling/importance_sampling_ratio/mean': 0.9963176250457764, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7998340651392937, 'clip_ratio/low_mean': 0.06532953213900328, 'clip_ratio/low_min': 0.06532953213900328, 'clip_ratio/high_mean': 0.04112681280821562, 'clip_ratio/high_max': 0.04112681280821562, 'clip_ratio/region_mean': 0.1064563449472189, 'reward_total_mean': 0.22415006160736084, 'reward_meter_mean': 0.4122018814086914, 'reward_meter_std': 0.45157957077026367, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.995192289352417, 'reward_lexical_plausibility_std': 0.013598216697573662, 'reward_repeat_penalty_mean': 0.9630240797996521, 'reward_repeat_penalty_std': 0.031217806041240692, 'reward_near_duplicate_penalty_mean': 0.9630240797996521, 'reward_near_duplicate_penalty_std': 0.031217806041240692, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.47999998927116394, 'reward_judge_quality_std': 0.28400203585624695, 'reward_total_composite_mean': 0.22415006160736084, 'reward_total_composite_std': 0.30988043546676636, 'epoch': 0.03}

 24%|██▍       | 784/3300 [3:06:57<6:31:18,  9.33s/it]INFO 04-12 13:53:09 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 785/3300 [3:07:05<6:21:56,  9.11s/it]2026-04-12 13:53:17,687 | INFO | train_grpo_train | metrics_logged mode=train step=785

                                                      
{'loss': 0.044, 'grad_norm': 12.129380226135254, 'learning_rate': 7.6242424242424254e-06, 'num_tokens': 1596896.0, 'completions/mean_length': 63.25, 'completions/min_length': 54.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.25, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.4240260124206543, 'rewards/meter/std': 0.26575201749801636, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.97265625, 'rewards/lexical_plausibility/std': 0.07733980566263199, 'rewards/judge_quality/mean': 0.8612500429153442, 'rewards/judge_quality/std': 0.16617010533809662, 'rewards/repeat_penalty/mean': 0.9878485202789307, 'rewards/repeat_penalty/std': 0.010732484050095081, 'rewards/near_duplicate_penalty/mean': 0.9878485202789307, 'rewards/near_duplicate_penalty/std': 0.010732484050095081, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3471768796443939, 'rewards/total_composite/std': 0.21641911566257477, 'reward': 0.3471768796443939, 'reward_std': 0.21641911566257477, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14089399576187134, 'sampling/sampling_logp_difference/max': 1.877039909362793, 'sampling/importance_sampling_ratio/min': 0.15304245054721832, 'sampling/importance_sampling_ratio/mean': 0.9881336688995361, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5656032934784889, 'clip_ratio/low_mean': 0.07480002893134952, 'clip_ratio/low_min': 0.07480002893134952, 'clip_ratio/high_mean': 0.021241830196231604, 'clip_ratio/high_max': 0.021241830196231604, 'clip_ratio/region_mean': 0.09604185912758112, 'reward_total_mean': 0.3471768796443939, 'reward_meter_mean': 0.4240260124206543, 'reward_meter_std': 0.26575201749801636, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.97265625, 'reward_lexical_plausibility_std': 0.07733980566263199, 'reward_repeat_penalty_mean': 0.9878485202789307, 'reward_repeat_penalty_std': 0.010732484050095081, 'reward_near_duplicate_penalty_mean': 0.9878485202789307, 'reward_near_duplicate_penalty_std': 0.010732484050095081, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8612500429153442, 'reward_judge_quality_std': 0.16617010533809662, 'reward_total_composite_mean': 0.3471768796443939, 'reward_total_composite_std': 0.21641911566257477, 'epoch': 0.03}

 24%|██▍       | 785/3300 [3:07:05<6:21:56,  9.11s/it]INFO 04-12 13:53:17 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 786/3300 [3:07:16<6:44:40,  9.66s/it]2026-04-12 13:53:28,617 | INFO | train_grpo_train | metrics_logged mode=train step=786

                                                      
{'loss': 0.0438, 'grad_norm': 9.026449203491211, 'learning_rate': 7.621212121212122e-06, 'num_tokens': 1600169.0, 'completions/mean_length': 194.125, 'completions/min_length': 158.0, 'completions/max_length': 220.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 194.125, 'completions/min_terminated_length': 158.0, 'completions/max_terminated_length': 220.0, 'rewards/meter/mean': 0.691949188709259, 'rewards/meter/std': 0.21175561845302582, 'rewards/count_adherence/mean': 0.9624999761581421, 'rewards/count_adherence/std': 0.05175493285059929, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.7400339841842651, 'rewards/repeat_penalty/std': 0.10134749114513397, 'rewards/near_duplicate_penalty/mean': 0.9417881369590759, 'rewards/near_duplicate_penalty/std': 0.02220279909670353, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.7936058044433594, 'rewards/distinct_2/std': 0.08217776566743851, 'rewards/total_composite/mean': 0.2555861473083496, 'rewards/total_composite/std': 0.07860279828310013, 'reward': 0.2555861473083496, 'reward_std': 0.07860280573368073, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14242398738861084, 'sampling/sampling_logp_difference/max': 3.973184108734131, 'sampling/importance_sampling_ratio/min': 0.01881343312561512, 'sampling/importance_sampling_ratio/mean': 0.9988805651664734, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6547017022967339, 'clip_ratio/low_mean': 0.05870208702981472, 'clip_ratio/low_min': 0.05870208702981472, 'clip_ratio/high_mean': 0.05752360261976719, 'clip_ratio/high_max': 0.05752360261976719, 'clip_ratio/region_mean': 0.11622568964958191, 'reward_total_mean': 0.2555861473083496, 'reward_meter_mean': 0.691949188709259, 'reward_meter_std': 0.21175561845302582, 'reward_count_adherence_mean': 0.9624999761581421, 'reward_count_adherence_std': 0.05175493285059929, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7400339841842651, 'reward_repeat_penalty_std': 0.10134749114513397, 'reward_near_duplicate_penalty_mean': 0.9417881369590759, 'reward_near_duplicate_penalty_std': 0.02220279909670353, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.7936058044433594, 'reward_distinct_2_std': 0.08217776566743851, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.2555861473083496, 'reward_total_composite_std': 0.07860279828310013, 'epoch': 0.03}

 24%|██▍       | 786/3300 [3:07:16<6:44:40,  9.66s/it]INFO 04-12 13:53:28 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 787/3300 [3:07:25<6:36:08,  9.46s/it]2026-04-12 13:53:37,610 | INFO | train_grpo_train | metrics_logged mode=train step=787

                                                      
{'loss': 0.0316, 'grad_norm': 9.279489517211914, 'learning_rate': 7.618181818181819e-06, 'num_tokens': 1602332.0, 'completions/mean_length': 91.375, 'completions/min_length': 72.0, 'completions/max_length': 112.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 91.375, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 112.0, 'rewards/meter/mean': 0.38759708404541016, 'rewards/meter/std': 0.31130579113960266, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5537499785423279, 'rewards/judge_quality/std': 0.22639647126197815, 'rewards/repeat_penalty/mean': 0.9654989242553711, 'rewards/repeat_penalty/std': 0.025990501046180725, 'rewards/near_duplicate_penalty/mean': 0.9810584783554077, 'rewards/near_duplicate_penalty/std': 0.009181052446365356, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9840800166130066, 'rewards/distinct_2/std': 0.022015659138560295, 'rewards/total_composite/mean': 0.21744802594184875, 'rewards/total_composite/std': 0.2141205370426178, 'reward': 0.21744802594184875, 'reward_std': 0.2141205370426178, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12602442502975464, 'sampling/sampling_logp_difference/max': 1.798635482788086, 'sampling/importance_sampling_ratio/min': 0.16552460193634033, 'sampling/importance_sampling_ratio/mean': 1.0107314586639404, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7529631331562996, 'clip_ratio/low_mean': 0.06275359215214849, 'clip_ratio/low_min': 0.06275359215214849, 'clip_ratio/high_mean': 0.050168365240097046, 'clip_ratio/high_max': 0.050168365240097046, 'clip_ratio/region_mean': 0.11292195739224553, 'reward_total_mean': 0.21744802594184875, 'reward_meter_mean': 0.38759708404541016, 'reward_meter_std': 0.31130579113960266, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.09258200973272324, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9654989242553711, 'reward_repeat_penalty_std': 0.025990501046180725, 'reward_near_duplicate_penalty_mean': 0.9810584783554077, 'reward_near_duplicate_penalty_std': 0.009181052446365356, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9840800166130066, 'reward_distinct_2_std': 0.022015659138560295, 'reward_judge_quality_mean': 0.5537499785423279, 'reward_judge_quality_std': 0.22639647126197815, 'reward_total_composite_mean': 0.21744802594184875, 'reward_total_composite_std': 0.2141205370426178, 'epoch': 0.03}

 24%|██▍       | 787/3300 [3:07:25<6:36:08,  9.46s/it]INFO 04-12 13:53:37 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 788/3300 [3:07:33<6:21:14,  9.11s/it]2026-04-12 13:53:45,893 | INFO | train_grpo_train | metrics_logged mode=train step=788

                                                      
{'loss': 0.078, 'grad_norm': 12.124133110046387, 'learning_rate': 7.6151515151515155e-06, 'num_tokens': 1604041.0, 'completions/mean_length': 54.625, 'completions/min_length': 49.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.625, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.6030672788619995, 'rewards/meter/std': 0.34272781014442444, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.48000001907348633, 'rewards/judge_quality/std': 0.3554876148700714, 'rewards/repeat_penalty/mean': 0.9607193470001221, 'rewards/repeat_penalty/std': 0.051950499415397644, 'rewards/near_duplicate_penalty/mean': 0.968349814414978, 'rewards/near_duplicate_penalty/std': 0.03550909087061882, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9916387796401978, 'rewards/distinct_2/std': 0.0236490610986948, 'rewards/total_composite/mean': 0.20431560277938843, 'rewards/total_composite/std': 0.12262222170829773, 'reward': 0.20431560277938843, 'reward_std': 0.12262222170829773, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13692446053028107, 'sampling/sampling_logp_difference/max': 1.3064236640930176, 'sampling/importance_sampling_ratio/min': 0.27078676223754883, 'sampling/importance_sampling_ratio/mean': 1.0283405780792236, 'sampling/importance_sampling_ratio/max': 1.9296555519104004, 'entropy': 1.1032139733433723, 'clip_ratio/low_mean': 0.0755410585552454, 'clip_ratio/low_min': 0.0755410585552454, 'clip_ratio/high_mean': 0.03683578735217452, 'clip_ratio/high_max': 0.03683578735217452, 'clip_ratio/region_mean': 0.11237684590741992, 'reward_total_mean': 0.20431560277938843, 'reward_meter_mean': 0.6030672788619995, 'reward_meter_std': 0.34272781014442444, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.9607193470001221, 'reward_repeat_penalty_std': 0.051950499415397644, 'reward_near_duplicate_penalty_mean': 0.968349814414978, 'reward_near_duplicate_penalty_std': 0.03550909087061882, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9916387796401978, 'reward_distinct_2_std': 0.0236490610986948, 'reward_judge_quality_mean': 0.48000001907348633, 'reward_judge_quality_std': 0.3554876148700714, 'reward_total_composite_mean': 0.20431560277938843, 'reward_total_composite_std': 0.12262222170829773, 'epoch': 0.03}

 24%|██▍       | 788/3300 [3:07:33<6:21:14,  9.11s/it]INFO 04-12 13:53:46 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 789/3300 [3:07:46<7:07:21, 10.21s/it]2026-04-12 13:53:58,691 | INFO | train_grpo_train | metrics_logged mode=train step=789

                                                      
{'loss': -0.0278, 'grad_norm': 2.4306137561798096, 'learning_rate': 7.612121212121213e-06, 'num_tokens': 1605424.0, 'completions/mean_length': 148.875, 'completions/min_length': 24.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 27.83333396911621, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.42209821939468384, 'rewards/meter/std': 0.46939533948898315, '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/lexical_plausibility/mean': 0.919418454170227, 'rewards/lexical_plausibility/std': 0.1502324342727661, 'rewards/judge_quality/mean': 0.45125001668930054, 'rewards/judge_quality/std': 0.40755149722099304, 'rewards/repeat_penalty/mean': 0.7149593234062195, 'rewards/repeat_penalty/std': 0.15700507164001465, 'rewards/near_duplicate_penalty/mean': 0.9116124510765076, 'rewards/near_duplicate_penalty/std': 0.14668512344360352, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1891888976097107, 'rewards/total_composite/std': 0.32353198528289795, 'reward': 0.1891888976097107, 'reward_std': 0.32353198528289795, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11858377605676651, 'sampling/sampling_logp_difference/max': 1.191333532333374, 'sampling/importance_sampling_ratio/min': 0.3038158416748047, 'sampling/importance_sampling_ratio/mean': 1.0269629955291748, 'sampling/importance_sampling_ratio/max': 1.8580149412155151, 'entropy': 0.6887340247631073, 'clip_ratio/low_mean': 0.0416162027977407, 'clip_ratio/low_min': 0.0416162027977407, 'clip_ratio/high_mean': 0.04620295576751232, 'clip_ratio/high_max': 0.04620295576751232, 'clip_ratio/region_mean': 0.08781915856525302, 'reward_total_mean': 0.1891888976097107, 'reward_meter_mean': 0.42209821939468384, 'reward_meter_std': 0.46939533948898315, '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_lexical_plausibility_mean': 0.919418454170227, 'reward_lexical_plausibility_std': 0.1502324342727661, 'reward_repeat_penalty_mean': 0.7149593234062195, 'reward_repeat_penalty_std': 0.15700507164001465, 'reward_near_duplicate_penalty_mean': 0.9116124510765076, 'reward_near_duplicate_penalty_std': 0.14668512344360352, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.45125001668930054, 'reward_judge_quality_std': 0.40755149722099304, 'reward_total_composite_mean': 0.1891888976097107, 'reward_total_composite_std': 0.32353198528289795, 'epoch': 0.03}

 24%|██▍       | 789/3300 [3:07:46<7:07:21, 10.21s/it]INFO 04-12 13:53:58 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 790/3300 [3:08:00<7:47:50, 11.18s/it]2026-04-12 13:54:12,137 | INFO | train_grpo_train | metrics_logged mode=train step=790

                                                      
{'loss': -0.0589, 'grad_norm': 5.683746337890625, 'learning_rate': 7.609090909090909e-06, 'num_tokens': 1607307.0, 'completions/mean_length': 122.375, 'completions/min_length': 51.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 66.71428680419922, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.4358466863632202, 'rewards/meter/std': 0.41290050745010376, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.17817415297031403, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9692785739898682, 'rewards/lexical_plausibility/std': 0.08689332008361816, 'rewards/judge_quality/mean': 0.3837500214576721, 'rewards/judge_quality/std': 0.2579555809497833, 'rewards/repeat_penalty/mean': 0.9191659688949585, 'rewards/repeat_penalty/std': 0.0840807631611824, 'rewards/near_duplicate_penalty/mean': 0.9589775800704956, 'rewards/near_duplicate_penalty/std': 0.029607703909277916, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9572770595550537, 'rewards/distinct_2/std': 0.06648191809654236, 'rewards/total_composite/mean': 0.14345701038837433, 'rewards/total_composite/std': 0.1522519737482071, 'reward': 0.14345701038837433, 'reward_std': 0.1522519588470459, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13306333124637604, 'sampling/sampling_logp_difference/max': 1.5143718719482422, 'sampling/importance_sampling_ratio/min': 0.21994629502296448, 'sampling/importance_sampling_ratio/mean': 1.016107201576233, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7931018397212029, 'clip_ratio/low_mean': 0.07569516729563475, 'clip_ratio/low_min': 0.07569516729563475, 'clip_ratio/high_mean': 0.03265977557748556, 'clip_ratio/high_max': 0.03265977557748556, 'clip_ratio/region_mean': 0.10835494287312031, 'reward_total_mean': 0.14345701038837433, 'reward_meter_mean': 0.4358466863632202, 'reward_meter_std': 0.41290050745010376, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.17817415297031403, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9692785739898682, 'reward_lexical_plausibility_std': 0.08689332008361816, 'reward_repeat_penalty_mean': 0.9191659688949585, 'reward_repeat_penalty_std': 0.0840807631611824, 'reward_near_duplicate_penalty_mean': 0.9589775800704956, 'reward_near_duplicate_penalty_std': 0.029607703909277916, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9572770595550537, 'reward_distinct_2_std': 0.06648191809654236, 'reward_judge_quality_mean': 0.3837500214576721, 'reward_judge_quality_std': 0.2579555809497833, 'reward_total_composite_mean': 0.14345701038837433, 'reward_total_composite_std': 0.1522519737482071, 'epoch': 0.03}

 24%|██▍       | 790/3300 [3:08:00<7:47:50, 11.18s/it]INFO 04-12 13:54:12 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 791/3300 [3:08:10<7:34:50, 10.88s/it]2026-04-12 13:54:22,298 | INFO | train_grpo_train | metrics_logged mode=train step=791

                                                      
{'loss': 0.0635, 'grad_norm': 6.78140115737915, 'learning_rate': 7.606060606060606e-06, 'num_tokens': 1610303.0, 'completions/mean_length': 173.5, 'completions/min_length': 152.0, 'completions/max_length': 197.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 173.5, 'completions/min_terminated_length': 152.0, 'completions/max_terminated_length': 197.0, 'rewards/meter/mean': 0.8277313709259033, 'rewards/meter/std': 0.1654488444328308, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.4860149025917053, 'rewards/repeat_penalty/std': 0.23334413766860962, 'rewards/near_duplicate_penalty/mean': 0.8298006057739258, 'rewards/near_duplicate_penalty/std': 0.06878306716680527, 'rewards/opening_diversity/mean': 0.984375, 'rewards/opening_diversity/std': 0.04419417306780815, 'rewards/distinct_2/mean': 0.6087020635604858, 'rewards/distinct_2/std': 0.2167627364397049, 'rewards/total_composite/mean': 0.18599867820739746, 'rewards/total_composite/std': 0.09120912849903107, 'reward': 0.18599867820739746, 'reward_std': 0.09120912849903107, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09823337942361832, 'sampling/sampling_logp_difference/max': 3.743842601776123, 'sampling/importance_sampling_ratio/min': 0.023663001134991646, 'sampling/importance_sampling_ratio/mean': 1.0083887577056885, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.579167727380991, 'clip_ratio/low_mean': 0.03915188554674387, 'clip_ratio/low_min': 0.03915188554674387, 'clip_ratio/high_mean': 0.046736656688153744, 'clip_ratio/high_max': 0.046736656688153744, 'clip_ratio/region_mean': 0.08588854223489761, 'reward_total_mean': 0.18599867820739746, 'reward_meter_mean': 0.8277313709259033, 'reward_meter_std': 0.1654488444328308, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.4860149025917053, 'reward_repeat_penalty_std': 0.23334413766860962, 'reward_near_duplicate_penalty_mean': 0.8298006057739258, 'reward_near_duplicate_penalty_std': 0.06878306716680527, 'reward_opening_diversity_mean': 0.984375, 'reward_opening_diversity_std': 0.04419417306780815, 'reward_distinct_2_mean': 0.6087020635604858, 'reward_distinct_2_std': 0.2167627364397049, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.18599867820739746, 'reward_total_composite_std': 0.09120912849903107, 'epoch': 0.03}

 24%|██▍       | 791/3300 [3:08:10<7:34:50, 10.88s/it]INFO 04-12 13:54:22 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 792/3300 [3:08:20<7:26:04, 10.67s/it]2026-04-12 13:54:32,491 | INFO | train_grpo_train | metrics_logged mode=train step=792

                                                      
{'loss': -0.0291, 'grad_norm': 7.867350101470947, 'learning_rate': 7.603030303030303e-06, 'num_tokens': 1612875.0, 'completions/mean_length': 117.5, 'completions/min_length': 99.0, 'completions/max_length': 136.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 117.5, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 136.0, 'rewards/meter/mean': 0.8110174536705017, 'rewards/meter/std': 0.2681174874305725, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.0353553369641304, 'rewards/repeat_penalty/mean': 0.6792880296707153, 'rewards/repeat_penalty/std': 0.12674009799957275, 'rewards/near_duplicate_penalty/mean': 0.8890895843505859, 'rewards/near_duplicate_penalty/std': 0.040204595774412155, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.7828903794288635, 'rewards/distinct_2/std': 0.0826951339840889, 'rewards/total_composite/mean': 0.2730887532234192, 'rewards/total_composite/std': 0.11727865040302277, 'reward': 0.2730887532234192, 'reward_std': 0.11727864295244217, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11835320293903351, 'sampling/sampling_logp_difference/max': 2.456092596054077, 'sampling/importance_sampling_ratio/min': 0.08576942980289459, 'sampling/importance_sampling_ratio/mean': 1.0024019479751587, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6475649625062943, 'clip_ratio/low_mean': 0.0405631959438324, 'clip_ratio/low_min': 0.0405631959438324, 'clip_ratio/high_mean': 0.0671022292226553, 'clip_ratio/high_max': 0.0671022292226553, 'clip_ratio/region_mean': 0.1076654251664877, 'reward_total_mean': 0.2730887532234192, 'reward_meter_mean': 0.8110174536705017, 'reward_meter_std': 0.2681174874305725, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6792880296707153, 'reward_repeat_penalty_std': 0.12674009799957275, 'reward_near_duplicate_penalty_mean': 0.8890895843505859, 'reward_near_duplicate_penalty_std': 0.040204595774412155, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.7828903794288635, 'reward_distinct_2_std': 0.0826951339840889, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.0353553369641304, 'reward_total_composite_mean': 0.2730887532234192, 'reward_total_composite_std': 0.11727865040302277, 'epoch': 0.03}

 24%|██▍       | 792/3300 [3:08:20<7:26:04, 10.67s/it]INFO 04-12 13:54:32 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 793/3300 [3:08:33<8:02:20, 11.54s/it]2026-04-12 13:54:46,073 | INFO | train_grpo_train | metrics_logged mode=train step=793

                                                      
{'loss': -0.0664, 'grad_norm': 2.1196558475494385, 'learning_rate': 7.600000000000001e-06, 'num_tokens': 1614471.0, 'completions/mean_length': 226.5, 'completions/min_length': 47.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 55.20000076293945, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.39400628209114075, 'rewards/meter/std': 0.43972936272621155, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.9326790571212769, 'rewards/lexical_plausibility/std': 0.09621042013168335, 'rewards/judge_quality/mean': 0.3212500214576721, 'rewards/judge_quality/std': 0.3066611886024475, 'rewards/repeat_penalty/mean': 0.9104909896850586, 'rewards/repeat_penalty/std': 0.1162523701786995, 'rewards/near_duplicate_penalty/mean': 0.9561881422996521, 'rewards/near_duplicate_penalty/std': 0.06250674277544022, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9825174808502197, 'rewards/distinct_2/std': 0.049448031932115555, 'rewards/total_composite/mean': 0.13824868202209473, 'rewards/total_composite/std': 0.17110635340213776, 'reward': 0.13824868202209473, 'reward_std': 0.17110635340213776, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13401292264461517, 'sampling/sampling_logp_difference/max': 1.46588134765625, 'sampling/importance_sampling_ratio/min': 0.2308744192123413, 'sampling/importance_sampling_ratio/mean': 1.0125057697296143, 'sampling/importance_sampling_ratio/max': 1.8767222166061401, 'entropy': 0.5378298386931419, 'clip_ratio/low_mean': 0.02801724150776863, 'clip_ratio/low_min': 0.02801724150776863, 'clip_ratio/high_mean': 0.050941724330186844, 'clip_ratio/high_max': 0.050941724330186844, 'clip_ratio/region_mean': 0.07895896583795547, 'reward_total_mean': 0.13824868202209473, 'reward_meter_mean': 0.39400628209114075, 'reward_meter_std': 0.43972936272621155, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.9326790571212769, 'reward_lexical_plausibility_std': 0.09621042013168335, 'reward_repeat_penalty_mean': 0.9104909896850586, 'reward_repeat_penalty_std': 0.1162523701786995, 'reward_near_duplicate_penalty_mean': 0.9561881422996521, 'reward_near_duplicate_penalty_std': 0.06250674277544022, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9825174808502197, 'reward_distinct_2_std': 0.049448031932115555, 'reward_judge_quality_mean': 0.3212500214576721, 'reward_judge_quality_std': 0.3066611886024475, 'reward_total_composite_mean': 0.13824868202209473, 'reward_total_composite_std': 0.17110635340213776, 'epoch': 0.03}

 24%|██▍       | 793/3300 [3:08:34<8:02:20, 11.54s/it]INFO 04-12 13:54:46 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 794/3300 [3:08:47<8:26:40, 12.13s/it]2026-04-12 13:54:59,570 | INFO | train_grpo_train | metrics_logged mode=train step=794

                                                      
{'loss': -0.0458, 'grad_norm': 1.9994498491287231, 'learning_rate': 7.596969696969697e-06, 'num_tokens': 1615952.0, 'completions/mean_length': 225.125, 'completions/min_length': 51.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 53.0, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.30815163254737854, 'rewards/meter/std': 0.4055858552455902, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.9001245498657227, 'rewards/lexical_plausibility/std': 0.13998135924339294, 'rewards/judge_quality/mean': 0.3462499976158142, 'rewards/judge_quality/std': 0.29875636100769043, 'rewards/repeat_penalty/mean': 0.8298996090888977, 'rewards/repeat_penalty/std': 0.17024771869182587, 'rewards/near_duplicate_penalty/mean': 0.9721620082855225, 'rewards/near_duplicate_penalty/std': 0.05000936985015869, 'rewards/opening_diversity/mean': 0.875, 'rewards/opening_diversity/std': 0.13363061845302582, 'rewards/distinct_2/mean': 0.9688255786895752, 'rewards/distinct_2/std': 0.06448943167924881, 'rewards/total_composite/mean': 0.08152542263269424, 'rewards/total_composite/std': 0.13616929948329926, 'reward': 0.08152542263269424, 'reward_std': 0.13616929948329926, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1460765153169632, 'sampling/sampling_logp_difference/max': 1.451857566833496, 'sampling/importance_sampling_ratio/min': 0.2341349720954895, 'sampling/importance_sampling_ratio/mean': 1.0301177501678467, 'sampling/importance_sampling_ratio/max': 1.8213194608688354, 'entropy': 0.8199094757437706, 'clip_ratio/low_mean': 0.04438250511884689, 'clip_ratio/low_min': 0.04438250511884689, 'clip_ratio/high_mean': 0.026405845768749714, 'clip_ratio/high_max': 0.026405845768749714, 'clip_ratio/region_mean': 0.07078835088759661, 'reward_total_mean': 0.08152542263269424, 'reward_meter_mean': 0.30815163254737854, 'reward_meter_std': 0.4055858552455902, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.9001245498657227, 'reward_lexical_plausibility_std': 0.13998135924339294, 'reward_repeat_penalty_mean': 0.8298996090888977, 'reward_repeat_penalty_std': 0.17024771869182587, 'reward_near_duplicate_penalty_mean': 0.9721620082855225, 'reward_near_duplicate_penalty_std': 0.05000936985015869, 'reward_opening_diversity_mean': 0.875, 'reward_opening_diversity_std': 0.13363061845302582, 'reward_distinct_2_mean': 0.9688255786895752, 'reward_distinct_2_std': 0.06448943167924881, 'reward_judge_quality_mean': 0.3462499976158142, 'reward_judge_quality_std': 0.29875636100769043, 'reward_total_composite_mean': 0.08152542263269424, 'reward_total_composite_std': 0.13616929948329926, 'epoch': 0.03}

 24%|██▍       | 794/3300 [3:08:47<8:26:40, 12.13s/it]INFO 04-12 13:54:59 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 795/3300 [3:08:56<7:46:41, 11.18s/it]2026-04-12 13:55:08,527 | INFO | train_grpo_train | metrics_logged mode=train step=795

                                                      
{'loss': 0.0811, 'grad_norm': 10.82763385772705, 'learning_rate': 7.593939393939395e-06, 'num_tokens': 1617655.0, 'completions/mean_length': 57.875, 'completions/min_length': 52.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.875, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.6357424855232239, 'rewards/meter/std': 0.36386656761169434, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.13093073666095734, 'rewards/repeat_penalty/mean': 0.9697063565254211, 'rewards/repeat_penalty/std': 0.04479420930147171, 'rewards/near_duplicate_penalty/mean': 0.9770572185516357, 'rewards/near_duplicate_penalty/std': 0.03484059497714043, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9923076629638672, 'rewards/distinct_2/std': 0.021757129579782486, 'rewards/total_composite/mean': 0.22125570476055145, 'rewards/total_composite/std': 0.1574130803346634, 'reward': 0.22125570476055145, 'reward_std': 0.1574130803346634, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14132018387317657, 'sampling/sampling_logp_difference/max': 1.8274240493774414, 'sampling/importance_sampling_ratio/min': 0.16082732379436493, 'sampling/importance_sampling_ratio/mean': 1.0085033178329468, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1949525251984596, 'clip_ratio/low_mean': 0.04530677758157253, 'clip_ratio/low_min': 0.04530677758157253, 'clip_ratio/high_mean': 0.06336268410086632, 'clip_ratio/high_max': 0.06336268410086632, 'clip_ratio/region_mean': 0.10866946168243885, 'reward_total_mean': 0.22125570476055145, 'reward_meter_mean': 0.6357424855232239, 'reward_meter_std': 0.36386656761169434, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9697063565254211, 'reward_repeat_penalty_std': 0.04479420930147171, 'reward_near_duplicate_penalty_mean': 0.9770572185516357, 'reward_near_duplicate_penalty_std': 0.03484059497714043, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9923076629638672, 'reward_distinct_2_std': 0.021757129579782486, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.13093073666095734, 'reward_total_composite_mean': 0.22125570476055145, 'reward_total_composite_std': 0.1574130803346634, 'epoch': 0.03}

 24%|██▍       | 795/3300 [3:08:56<7:46:41, 11.18s/it]INFO 04-12 13:55:08 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 796/3300 [3:09:11<8:35:12, 12.35s/it]2026-04-12 13:55:23,596 | INFO | train_grpo_train | metrics_logged mode=train step=796

                                                      
{'loss': -0.1423, 'grad_norm': 3.7218844890594482, 'learning_rate': 7.590909090909091e-06, 'num_tokens': 1620466.0, 'completions/mean_length': 227.375, 'completions/min_length': 165.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 186.71429443359375, 'completions/min_terminated_length': 165.0, 'completions/max_terminated_length': 212.0, 'rewards/meter/mean': 0.8557415008544922, 'rewards/meter/std': 0.2478206753730774, 'rewards/count_adherence/mean': 0.8035714626312256, 'rewards/count_adherence/std': 0.15152288973331451, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9750967621803284, 'rewards/lexical_plausibility/std': 0.070436991751194, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.647225022315979, 'rewards/repeat_penalty/std': 0.1830524355173111, 'rewards/near_duplicate_penalty/mean': 0.9145234227180481, 'rewards/near_duplicate_penalty/std': 0.04553912580013275, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.737200140953064, 'rewards/distinct_2/std': 0.13940955698490143, 'rewards/total_composite/mean': 0.19946572184562683, 'rewards/total_composite/std': 0.11524349451065063, 'reward': 0.19946572184562683, 'reward_std': 0.11524349451065063, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11161487549543381, 'sampling/sampling_logp_difference/max': 1.481365442276001, 'sampling/importance_sampling_ratio/min': 0.22732707858085632, 'sampling/importance_sampling_ratio/mean': 1.0250014066696167, 'sampling/importance_sampling_ratio/max': 1.9287081956863403, 'entropy': 0.9043225273489952, 'clip_ratio/low_mean': 0.021276067942380905, 'clip_ratio/low_min': 0.021276067942380905, 'clip_ratio/high_mean': 0.06665316876024008, 'clip_ratio/high_max': 0.06665316876024008, 'clip_ratio/region_mean': 0.08792923670262098, 'reward_total_mean': 0.19946572184562683, 'reward_meter_mean': 0.8557415008544922, 'reward_meter_std': 0.2478206753730774, 'reward_count_adherence_mean': 0.8035714626312256, 'reward_count_adherence_std': 0.15152288973331451, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9750967621803284, 'reward_lexical_plausibility_std': 0.070436991751194, 'reward_repeat_penalty_mean': 0.647225022315979, 'reward_repeat_penalty_std': 0.1830524355173111, 'reward_near_duplicate_penalty_mean': 0.9145234227180481, 'reward_near_duplicate_penalty_std': 0.04553912580013275, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.737200140953064, 'reward_distinct_2_std': 0.13940955698490143, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.19946572184562683, 'reward_total_composite_std': 0.11524349451065063, 'epoch': 0.03}

 24%|██▍       | 796/3300 [3:09:11<8:35:12, 12.35s/it]INFO 04-12 13:55:23 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 797/3300 [3:09:25<8:53:34, 12.79s/it]2026-04-12 13:55:37,426 | INFO | train_grpo_train | metrics_logged mode=train step=797

                                                      
{'loss': -0.0865, 'grad_norm': 2.3491971492767334, 'learning_rate': 7.587878787878788e-06, 'num_tokens': 1622248.0, 'completions/mean_length': 231.75, 'completions/min_length': 55.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 63.60000228881836, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.8426979184150696, 'rewards/meter/std': 0.2519124448299408, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9274377822875977, 'rewards/lexical_plausibility/std': 0.08428309857845306, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.19086270034313202, 'rewards/repeat_penalty/mean': 0.6938273310661316, 'rewards/repeat_penalty/std': 0.20207174122333527, 'rewards/near_duplicate_penalty/mean': 0.894913375377655, 'rewards/near_duplicate_penalty/std': 0.09684357792139053, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9376318454742432, 'rewards/distinct_2/std': 0.08143006265163422, 'rewards/total_composite/mean': 0.2021869719028473, 'rewards/total_composite/std': 0.1991923749446869, 'reward': 0.2021869719028473, 'reward_std': 0.1991923600435257, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12575359642505646, 'sampling/sampling_logp_difference/max': 1.5154848098754883, 'sampling/importance_sampling_ratio/min': 0.2197016477584839, 'sampling/importance_sampling_ratio/mean': 1.016856074333191, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6785863637924194, 'clip_ratio/low_mean': 0.04193564131855965, 'clip_ratio/low_min': 0.04193564131855965, 'clip_ratio/high_mean': 0.03797266911715269, 'clip_ratio/high_max': 0.03797266911715269, 'clip_ratio/region_mean': 0.07990831043571234, 'reward_total_mean': 0.2021869719028473, 'reward_meter_mean': 0.8426979184150696, 'reward_meter_std': 0.2519124448299408, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9274377822875977, 'reward_lexical_plausibility_std': 0.08428309857845306, 'reward_repeat_penalty_mean': 0.6938273310661316, 'reward_repeat_penalty_std': 0.20207174122333527, 'reward_near_duplicate_penalty_mean': 0.894913375377655, 'reward_near_duplicate_penalty_std': 0.09684357792139053, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9376318454742432, 'reward_distinct_2_std': 0.08143006265163422, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.19086270034313202, 'reward_total_composite_mean': 0.2021869719028473, 'reward_total_composite_std': 0.1991923749446869, 'epoch': 0.03}

 24%|██▍       | 797/3300 [3:09:25<8:53:34, 12.79s/it]INFO 04-12 13:55:37 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 798/3300 [3:09:39<9:08:24, 13.15s/it]2026-04-12 13:55:51,420 | INFO | train_grpo_train | metrics_logged mode=train step=798

                                                      
{'loss': -0.045, 'grad_norm': 4.649588108062744, 'learning_rate': 7.584848484848486e-06, 'num_tokens': 1623895.0, 'completions/mean_length': 106.875, 'completions/min_length': 41.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 49.000003814697266, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.6023092865943909, 'rewards/meter/std': 0.4553332030773163, '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/lexical_plausibility/mean': 0.9571899175643921, 'rewards/lexical_plausibility/std': 0.1210852637887001, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.9110137820243835, 'rewards/repeat_penalty/std': 0.09770961850881577, 'rewards/near_duplicate_penalty/mean': 0.959513783454895, 'rewards/near_duplicate_penalty/std': 0.04790858179330826, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9807209968566895, 'rewards/distinct_2/std': 0.0357571542263031, 'rewards/total_composite/mean': 0.1801133155822754, 'rewards/total_composite/std': 0.16981984674930573, 'reward': 0.1801133155822754, 'reward_std': 0.16981984674930573, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14103667438030243, 'sampling/sampling_logp_difference/max': 1.1155471801757812, 'sampling/importance_sampling_ratio/min': 0.32773590087890625, 'sampling/importance_sampling_ratio/mean': 1.0315786600112915, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1726729571819305, 'clip_ratio/low_mean': 0.08622794505208731, 'clip_ratio/low_min': 0.08622794505208731, 'clip_ratio/high_mean': 0.04975543078035116, 'clip_ratio/high_max': 0.04975543078035116, 'clip_ratio/region_mean': 0.13598337583243847, 'reward_total_mean': 0.1801133155822754, 'reward_meter_mean': 0.6023092865943909, 'reward_meter_std': 0.4553332030773163, '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_lexical_plausibility_mean': 0.9571899175643921, 'reward_lexical_plausibility_std': 0.1210852637887001, 'reward_repeat_penalty_mean': 0.9110137820243835, 'reward_repeat_penalty_std': 0.09770961850881577, 'reward_near_duplicate_penalty_mean': 0.959513783454895, 'reward_near_duplicate_penalty_std': 0.04790858179330826, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9807209968566895, 'reward_distinct_2_std': 0.0357571542263031, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.1801133155822754, 'reward_total_composite_std': 0.16981984674930573, 'epoch': 0.03}

 24%|██▍       | 798/3300 [3:09:39<9:08:24, 13.15s/it]INFO 04-12 13:55:51 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 799/3300 [3:09:53<9:17:55, 13.38s/it]2026-04-12 13:56:05,346 | INFO | train_grpo_train | metrics_logged mode=train step=799

                                                      
{'loss': -0.2154, 'grad_norm': 2.5473053455352783, 'learning_rate': 7.581818181818183e-06, 'num_tokens': 1626541.0, 'completions/mean_length': 260.75, 'completions/min_length': 160.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 177.0, 'completions/min_terminated_length': 160.0, 'completions/max_terminated_length': 188.0, 'rewards/meter/mean': 0.3980942964553833, 'rewards/meter/std': 0.3225463926792145, 'rewards/count_adherence/mean': 0.7142857313156128, 'rewards/count_adherence/std': 0.27532118558883667, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9650053381919861, 'rewards/lexical_plausibility/std': 0.06587570905685425, 'rewards/judge_quality/mean': 0.17499999701976776, 'rewards/judge_quality/std': 0.11649647355079651, 'rewards/repeat_penalty/mean': 0.6902345418930054, 'rewards/repeat_penalty/std': 0.24288712441921234, 'rewards/near_duplicate_penalty/mean': 0.9081143736839294, 'rewards/near_duplicate_penalty/std': 0.06450329720973969, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0578637570142746, 'rewards/distinct_2/mean': 0.7805726528167725, 'rewards/distinct_2/std': 0.18296225368976593, 'rewards/total_composite/mean': 0.06317733228206635, 'rewards/total_composite/std': 0.04797625169157982, 'reward': 0.06317733228206635, 'reward_std': 0.04797625169157982, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12680472433567047, 'sampling/sampling_logp_difference/max': 1.7284345626831055, 'sampling/importance_sampling_ratio/min': 0.1775621473789215, 'sampling/importance_sampling_ratio/mean': 1.0184462070465088, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8330177888274193, 'clip_ratio/low_mean': 0.025344843044877052, 'clip_ratio/low_min': 0.025344843044877052, 'clip_ratio/high_mean': 0.06401803530752659, 'clip_ratio/high_max': 0.06401803530752659, 'clip_ratio/region_mean': 0.08936287835240364, 'reward_total_mean': 0.06317733228206635, 'reward_meter_mean': 0.3980942964553833, 'reward_meter_std': 0.3225463926792145, 'reward_count_adherence_mean': 0.7142857313156128, 'reward_count_adherence_std': 0.27532118558883667, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9650053381919861, 'reward_lexical_plausibility_std': 0.06587570905685425, 'reward_repeat_penalty_mean': 0.6902345418930054, 'reward_repeat_penalty_std': 0.24288712441921234, 'reward_near_duplicate_penalty_mean': 0.9081143736839294, 'reward_near_duplicate_penalty_std': 0.06450329720973969, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0578637570142746, 'reward_distinct_2_mean': 0.7805726528167725, 'reward_distinct_2_std': 0.18296225368976593, 'reward_judge_quality_mean': 0.17499999701976776, 'reward_judge_quality_std': 0.11649647355079651, 'reward_total_composite_mean': 0.06317733228206635, 'reward_total_composite_std': 0.04797625169157982, 'epoch': 0.03}

 24%|██▍       | 799/3300 [3:09:53<9:17:55, 13.38s/it]INFO 04-12 13:56:05 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 800/3300 [3:10:07<9:26:32, 13.60s/it]2026-04-12 13:56:19,439 | INFO | train_grpo_train | metrics_logged mode=train step=800

                                                      
{'loss': -0.0944, 'grad_norm': 2.552888870239258, 'learning_rate': 7.57878787878788e-06, 'num_tokens': 1628227.0, 'completions/mean_length': 242.75, 'completions/min_length': 76.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 81.20000457763672, 'completions/min_terminated_length': 76.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.446036696434021, 'rewards/meter/std': 0.38267219066619873, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.30860668420791626, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.9024382829666138, 'rewards/lexical_plausibility/std': 0.134896382689476, 'rewards/judge_quality/mean': 0.21250000596046448, 'rewards/judge_quality/std': 0.1505940705537796, 'rewards/repeat_penalty/mean': 0.7273341417312622, 'rewards/repeat_penalty/std': 0.23968623578548431, 'rewards/near_duplicate_penalty/mean': 0.9229803085327148, 'rewards/near_duplicate_penalty/std': 0.06741784512996674, 'rewards/opening_diversity/mean': 0.84375, 'rewards/opening_diversity/std': 0.18600596487522125, 'rewards/distinct_2/mean': 0.909904956817627, 'rewards/distinct_2/std': 0.08924520760774612, 'rewards/total_composite/mean': 0.09542323648929596, 'rewards/total_composite/std': 0.11714858561754227, 'reward': 0.09542323648929596, 'reward_std': 0.11714857816696167, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11899919807910919, 'sampling/sampling_logp_difference/max': 1.0878639221191406, 'sampling/importance_sampling_ratio/min': 0.33693546056747437, 'sampling/importance_sampling_ratio/mean': 1.0266708135604858, 'sampling/importance_sampling_ratio/max': 1.9578245878219604, 'entropy': 0.521456390619278, 'clip_ratio/low_mean': 0.035671756602823734, 'clip_ratio/low_min': 0.035671756602823734, 'clip_ratio/high_mean': 0.04763440228998661, 'clip_ratio/high_max': 0.04763440228998661, 'clip_ratio/region_mean': 0.08330615889281034, 'reward_total_mean': 0.09542323648929596, 'reward_meter_mean': 0.446036696434021, 'reward_meter_std': 0.38267219066619873, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.30860668420791626, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.9024382829666138, 'reward_lexical_plausibility_std': 0.134896382689476, 'reward_repeat_penalty_mean': 0.7273341417312622, 'reward_repeat_penalty_std': 0.23968623578548431, 'reward_near_duplicate_penalty_mean': 0.9229803085327148, 'reward_near_duplicate_penalty_std': 0.06741784512996674, 'reward_opening_diversity_mean': 0.84375, 'reward_opening_diversity_std': 0.18600596487522125, 'reward_distinct_2_mean': 0.909904956817627, 'reward_distinct_2_std': 0.08924520760774612, 'reward_judge_quality_mean': 0.21250000596046448, 'reward_judge_quality_std': 0.1505940705537796, 'reward_total_composite_mean': 0.09542323648929596, 'reward_total_composite_std': 0.11714858561754227, 'epoch': 0.03}

 24%|██▍       | 800/3300 [3:10:07<9:26:32, 13.60s/it]INFO 04-12 13:56:19 [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 [00:11<01:02,  5.64s/it]

 23%|██▎       | 3/13 [00:22<01:19,  7.95s/it]

 31%|███       | 4/13 [00:34<01:23,  9.32s/it]

 38%|███▊      | 5/13 [00:45<01:19,  9.90s/it]

 46%|████▌     | 6/13 [00:56<01:13, 10.46s/it]

 54%|█████▍    | 7/13 [01:07<01:03, 10.66s/it]

 62%|██████▏   | 8/13 [01:19<00:55, 11.08s/it]

 69%|██████▉   | 9/13 [01:31<00:44, 11.22s/it]

 77%|███████▋  | 10/13 [01:42<00:33, 11.15s/it]

 85%|████████▍ | 11/13 [01:53<00:22, 11.22s/it]

 92%|█████████▏| 12/13 [02:04<00:11, 11.15s/it]

100%|██████████| 13/13 [02:16<00:00, 11.35s/it]2026-04-12 13:58:47,753 | INFO | train_grpo_train | metrics_logged mode=eval step=800

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 148.2555, 'eval_samples_per_second': 0.701, 'eval_steps_per_second': 0.088, 'eval_num_tokens': 1628227.0, 'eval_completions/mean_length': 251.14423076923077, 'eval_completions/min_length': 53.30769230769231, 'eval_completions/max_length': 512.0, 'eval_completions/clipped_ratio': 0.25961538461538464, 'eval_completions/mean_terminated_length': 158.62601118821365, 'eval_completions/min_terminated_length': 53.30769230769231, 'eval_completions/max_terminated_length': 304.53846153846155, 'eval_rewards/meter/mean': 0.509127662732051, 'eval_rewards/meter/std': 0.36145220696926117, 'eval_rewards/count_adherence/mean': 0.8326160632647001, 'eval_rewards/count_adherence/std': 0.21441587977684462, 'eval_rewards/arabic_clean/mean': 0.8173076923076923, 'eval_rewards/arabic_clean/std': 0.31626112644488996, 'eval_rewards/lexical_plausibility/mean': 0.958170601954827, 'eval_rewards/lexical_plausibility/std': 0.08274339003345141, 'eval_rewards/judge_quality/mean': 0.30413461648500884, 'eval_rewards/judge_quality/std': 0.1935537434541262, 'eval_rewards/repeat_penalty/mean': 0.6384293872576493, 'eval_rewards/repeat_penalty/std': 0.3749214456631587, 'eval_rewards/near_duplicate_penalty/mean': 0.8722778191933265, 'eval_rewards/near_duplicate_penalty/std': 0.1579323072846119, 'eval_rewards/opening_diversity/mean': 0.9181619240687444, 'eval_rewards/opening_diversity/std': 0.16182654408308175, 'eval_rewards/distinct_2/mean': 0.7001253549869244, 'eval_rewards/distinct_2/std': 0.3752513281427897, 'eval_rewards/total_composite/mean': 0.142544326873926, 'eval_rewards/total_composite/std': 0.1466588183091237, 'eval_reward': 0.142544326873926, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.0514264189853118, 'eval_sampling/sampling_logp_difference/max': 0.9607755220853366, 'eval_sampling/importance_sampling_ratio/min': 0.3911426869722513, 'eval_sampling/importance_sampling_ratio/mean': 1.0133316058378954, 'eval_sampling/importance_sampling_ratio/max': 1.4807185668211718, 'eval_entropy': 0.6052966370032384, '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': 0.142544326873926, 'eval_reward_meter_mean': 0.509127662732051, 'eval_reward_meter_std': 0.36145220696926117, 'eval_reward_count_adherence_mean': 0.8326160632647001, 'eval_reward_count_adherence_std': 0.21441587977684462, 'eval_reward_arabic_clean_mean': 0.8173076923076923, 'eval_reward_arabic_clean_std': 0.31626112644488996, 'eval_reward_lexical_plausibility_mean': 0.958170601954827, 'eval_reward_lexical_plausibility_std': 0.08274339003345141, 'eval_reward_repeat_penalty_mean': 0.6384293872576493, 'eval_reward_repeat_penalty_std': 0.3749214456631587, 'eval_reward_near_duplicate_penalty_mean': 0.8722778191933265, 'eval_reward_near_duplicate_penalty_std': 0.1579323072846119, 'eval_reward_opening_diversity_mean': 0.9181619240687444, 'eval_reward_opening_diversity_std': 0.16182654408308175, 'eval_reward_distinct_2_mean': 0.7001253549869244, 'eval_reward_distinct_2_std': 0.3752513281427897, 'eval_reward_judge_quality_mean': 0.30413461648500884, 'eval_reward_judge_quality_std': 0.1935537434541262, 'eval_reward_total_composite_mean': 0.142544326873926, 'eval_reward_total_composite_std': 0.1466588183091237, 'epoch': 0.03}

 24%|██▍       | 800/3300 [3:12:35<9:26:32, 13.60s/it]

100%|██████████| 13/13 [02:16<00:00, 11.35s/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 [3:12:51<40:48:16, 58.78s/it]2026-04-12 13:59:03,656 | INFO | train_grpo_train | metrics_logged mode=train step=801

                                                       
{'loss': -0.0903, 'grad_norm': 4.7603678703308105, 'learning_rate': 7.5757575757575764e-06, 'num_tokens': 1629797.0, 'completions/mean_length': 104.25, 'completions/min_length': 41.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 46.000003814697266, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.8133636116981506, 'rewards/meter/std': 0.307601660490036, '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/lexical_plausibility/mean': 0.9706271290779114, 'rewards/lexical_plausibility/std': 0.0830790102481842, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.16035674512386322, 'rewards/repeat_penalty/mean': 0.8802578449249268, 'rewards/repeat_penalty/std': 0.09642595052719116, 'rewards/near_duplicate_penalty/mean': 0.9326709508895874, 'rewards/near_duplicate_penalty/std': 0.03488456457853317, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.942307710647583, 'rewards/distinct_2/std': 0.07962294667959213, 'rewards/total_composite/mean': 0.3017697334289551, 'rewards/total_composite/std': 0.18662135303020477, 'reward': 0.3017697334289551, 'reward_std': 0.18662135303020477, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15090008080005646, 'sampling/sampling_logp_difference/max': 1.8050336837768555, 'sampling/importance_sampling_ratio/min': 0.164468914270401, 'sampling/importance_sampling_ratio/mean': 1.0130020380020142, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8559260070323944, 'clip_ratio/low_mean': 0.015243902802467346, 'clip_ratio/low_min': 0.015243902802467346, 'clip_ratio/high_mean': 0.09896360617130995, 'clip_ratio/high_max': 0.09896360617130995, 'clip_ratio/region_mean': 0.1142075089737773, 'reward_total_mean': 0.3017697334289551, 'reward_meter_mean': 0.8133636116981506, 'reward_meter_std': 0.307601660490036, '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_lexical_plausibility_mean': 0.9706271290779114, 'reward_lexical_plausibility_std': 0.0830790102481842, 'reward_repeat_penalty_mean': 0.8802578449249268, 'reward_repeat_penalty_std': 0.09642595052719116, 'reward_near_duplicate_penalty_mean': 0.9326709508895874, 'reward_near_duplicate_penalty_std': 0.03488456457853317, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.942307710647583, 'reward_distinct_2_std': 0.07962294667959213, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.16035674512386322, 'reward_total_composite_mean': 0.3017697334289551, 'reward_total_composite_std': 0.18662135303020477, 'epoch': 0.03}

 24%|██▍       | 801/3300 [3:12:51<40:48:16, 58.78s/it]INFO 04-12 13:59:03 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 802/3300 [3:13:04<31:09:49, 44.91s/it]2026-04-12 13:59:16,203 | INFO | train_grpo_train | metrics_logged mode=train step=802

                                                       
{'loss': -0.0453, 'grad_norm': 3.8556666374206543, 'learning_rate': 7.572727272727274e-06, 'num_tokens': 1631230.0, 'completions/mean_length': 87.125, 'completions/min_length': 24.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 26.428571701049805, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.7231729030609131, 'rewards/meter/std': 0.4302844703197479, '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/lexical_plausibility/mean': 0.9578319787979126, 'rewards/lexical_plausibility/std': 0.07433262467384338, 'rewards/judge_quality/mean': 0.71875, 'rewards/judge_quality/std': 0.3829747438430786, 'rewards/repeat_penalty/mean': 0.7019463181495667, 'rewards/repeat_penalty/std': 0.07367491722106934, 'rewards/near_duplicate_penalty/mean': 0.935928463935852, 'rewards/near_duplicate_penalty/std': 0.09823321551084518, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5763436555862427, 'rewards/total_composite/std': 0.4322705566883087, 'reward': 0.5763436555862427, 'reward_std': 0.43227052688598633, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12684661149978638, 'sampling/sampling_logp_difference/max': 1.168478012084961, 'sampling/importance_sampling_ratio/min': 0.3108396530151367, 'sampling/importance_sampling_ratio/mean': 1.0102213621139526, 'sampling/importance_sampling_ratio/max': 1.6923881769180298, 'entropy': 0.7740373127162457, 'clip_ratio/low_mean': 0.024193547666072845, 'clip_ratio/low_min': 0.024193547666072845, 'clip_ratio/high_mean': 0.0625854698009789, 'clip_ratio/high_max': 0.0625854698009789, 'clip_ratio/region_mean': 0.08677901746705174, 'reward_total_mean': 0.5763436555862427, 'reward_meter_mean': 0.7231729030609131, 'reward_meter_std': 0.4302844703197479, '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_lexical_plausibility_mean': 0.9578319787979126, 'reward_lexical_plausibility_std': 0.07433262467384338, 'reward_repeat_penalty_mean': 0.7019463181495667, 'reward_repeat_penalty_std': 0.07367491722106934, 'reward_near_duplicate_penalty_mean': 0.935928463935852, 'reward_near_duplicate_penalty_std': 0.09823321551084518, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.71875, 'reward_judge_quality_std': 0.3829747438430786, 'reward_total_composite_mean': 0.5763436555862427, 'reward_total_composite_std': 0.4322705566883087, 'epoch': 0.03}

 24%|██▍       | 802/3300 [3:13:04<31:09:49, 44.91s/it]INFO 04-12 13:59:16 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 803/3300 [3:13:17<24:29:20, 35.31s/it]2026-04-12 13:59:29,097 | INFO | train_grpo_train | metrics_logged mode=train step=803

                                                       
{'loss': -0.0488, 'grad_norm': 1.8177719116210938, 'learning_rate': 7.56969696969697e-06, 'num_tokens': 1632873.0, 'completions/mean_length': 222.375, 'completions/min_length': 44.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 48.60000228881836, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.45540738105773926, 'rewards/meter/std': 0.4550536572933197, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.9265633821487427, 'rewards/lexical_plausibility/std': 0.10626909136772156, 'rewards/judge_quality/mean': 0.1875, 'rewards/judge_quality/std': 0.14078859984874725, 'rewards/repeat_penalty/mean': 0.7691882848739624, 'rewards/repeat_penalty/std': 0.2682081460952759, 'rewards/near_duplicate_penalty/mean': 0.9177781343460083, 'rewards/near_duplicate_penalty/std': 0.10729513317346573, 'rewards/opening_diversity/mean': 0.925000011920929, 'rewards/opening_diversity/std': 0.11338934302330017, 'rewards/distinct_2/mean': 0.9015784859657288, 'rewards/distinct_2/std': 0.15883569419384003, 'rewards/total_composite/mean': 0.09251372516155243, 'rewards/total_composite/std': 0.1316995769739151, 'reward': 0.09251372516155243, 'reward_std': 0.1316995769739151, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13744182884693146, 'sampling/sampling_logp_difference/max': 2.3321428298950195, 'sampling/importance_sampling_ratio/min': 0.09708748012781143, 'sampling/importance_sampling_ratio/mean': 1.013674020767212, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5119334906339645, 'clip_ratio/low_mean': 0.03316326532512903, 'clip_ratio/low_min': 0.03316326532512903, 'clip_ratio/high_mean': 0.04126336984336376, 'clip_ratio/high_max': 0.04126336984336376, 'clip_ratio/region_mean': 0.0744266351684928, 'reward_total_mean': 0.09251372516155243, 'reward_meter_mean': 0.45540738105773926, 'reward_meter_std': 0.4550536572933197, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.9265633821487427, 'reward_lexical_plausibility_std': 0.10626909136772156, 'reward_repeat_penalty_mean': 0.7691882848739624, 'reward_repeat_penalty_std': 0.2682081460952759, 'reward_near_duplicate_penalty_mean': 0.9177781343460083, 'reward_near_duplicate_penalty_std': 0.10729513317346573, 'reward_opening_diversity_mean': 0.925000011920929, 'reward_opening_diversity_std': 0.11338934302330017, 'reward_distinct_2_mean': 0.9015784859657288, 'reward_distinct_2_std': 0.15883569419384003, 'reward_judge_quality_mean': 0.1875, 'reward_judge_quality_std': 0.14078859984874725, 'reward_total_composite_mean': 0.09251372516155243, 'reward_total_composite_std': 0.1316995769739151, 'epoch': 0.03}

 24%|██▍       | 803/3300 [3:13:17<24:29:20, 35.31s/it]INFO 04-12 13:59:29 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 804/3300 [3:13:30<20:02:05, 28.90s/it]2026-04-12 13:59:43,035 | INFO | train_grpo_train | metrics_logged mode=train step=804

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 7.566666666666667e-06, 'num_tokens': 1634537.0, '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': 0.8814504146575928, 'rewards/meter/std': 0.14642086625099182, 'rewards/count_adherence/mean': 0.824999988079071, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 0.6114320158958435, 'rewards/near_duplicate_penalty/std': 0.0560067854821682, 'rewards/opening_diversity/mean': 0.83519446849823, 'rewards/opening_diversity/std': 0.08232733607292175, 'rewards/distinct_2/mean': 0.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1087028980255127, 'rewards/total_composite/std': 0.018740905448794365, 'reward': 0.1087028980255127, 'reward_std': 0.018740905448794365, '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': 0.1087028980255127, 'reward_meter_mean': 0.8814504146575928, 'reward_meter_std': 0.14642086625099182, 'reward_count_adherence_mean': 0.824999988079071, 'reward_count_adherence_std': 0.0707106739282608, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.0, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 0.6114320158958435, 'reward_near_duplicate_penalty_std': 0.0560067854821682, 'reward_opening_diversity_mean': 0.83519446849823, 'reward_opening_diversity_std': 0.08232733607292175, 'reward_distinct_2_mean': 0.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.1087028980255127, 'reward_total_composite_std': 0.018740905448794365, 'epoch': 0.03}

 24%|██▍       | 804/3300 [3:13:31<20:02:05, 28.90s/it]INFO 04-12 13:59:43 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 805/3300 [3:13:43<16:42:38, 24.11s/it]2026-04-12 13:59:55,986 | INFO | train_grpo_train | metrics_logged mode=train step=805

                                                       
{'loss': -0.0881, 'grad_norm': 3.9951589107513428, 'learning_rate': 7.563636363636364e-06, 'num_tokens': 1636190.0, 'completions/mean_length': 111.625, 'completions/min_length': 50.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 54.42857360839844, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.29652339220046997, 'rewards/meter/std': 0.25362810492515564, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9670965671539307, 'rewards/lexical_plausibility/std': 0.09306494146585464, 'rewards/judge_quality/mean': 0.45625001192092896, 'rewards/judge_quality/std': 0.2859039604663849, 'rewards/repeat_penalty/mean': 0.9841243028640747, 'rewards/repeat_penalty/std': 0.023894984275102615, 'rewards/near_duplicate_penalty/mean': 0.9841243028640747, 'rewards/near_duplicate_penalty/std': 0.023894984275102615, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1305471658706665, 'rewards/total_composite/std': 0.09607304632663727, 'reward': 0.1305471658706665, 'reward_std': 0.09607304632663727, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12095951288938522, 'sampling/sampling_logp_difference/max': 1.6738252639770508, 'sampling/importance_sampling_ratio/min': 0.1875283420085907, 'sampling/importance_sampling_ratio/mean': 1.0074149370193481, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6449631676077843, 'clip_ratio/low_mean': 0.045387988910079, 'clip_ratio/low_min': 0.045387988910079, 'clip_ratio/high_mean': 0.062398978509008884, 'clip_ratio/high_max': 0.062398978509008884, 'clip_ratio/region_mean': 0.10778696741908789, 'reward_total_mean': 0.1305471658706665, 'reward_meter_mean': 0.29652339220046997, 'reward_meter_std': 0.25362810492515564, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9670965671539307, 'reward_lexical_plausibility_std': 0.09306494146585464, 'reward_repeat_penalty_mean': 0.9841243028640747, 'reward_repeat_penalty_std': 0.023894984275102615, 'reward_near_duplicate_penalty_mean': 0.9841243028640747, 'reward_near_duplicate_penalty_std': 0.023894984275102615, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.45625001192092896, 'reward_judge_quality_std': 0.2859039604663849, 'reward_total_composite_mean': 0.1305471658706665, 'reward_total_composite_std': 0.09607304632663727, 'epoch': 0.03}

 24%|██▍       | 805/3300 [3:13:43<16:42:38, 24.11s/it]INFO 04-12 13:59:56 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 806/3300 [3:13:52<13:32:08, 19.54s/it]2026-04-12 14:00:04,846 | INFO | train_grpo_train | metrics_logged mode=train step=806

                                                       
{'loss': 0.0586, 'grad_norm': 9.549102783203125, 'learning_rate': 7.560606060606062e-06, 'num_tokens': 1637925.0, 'completions/mean_length': 52.875, 'completions/min_length': 48.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.875, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.27364253997802734, 'rewards/meter/std': 0.4285583198070526, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.7717296481132507, 'rewards/repeat_penalty/std': 0.24070711433887482, 'rewards/near_duplicate_penalty/mean': 0.9155372381210327, 'rewards/near_duplicate_penalty/std': 0.057281166315078735, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9046062231063843, 'rewards/distinct_2/std': 0.15360285341739655, 'rewards/total_composite/mean': 0.10954989492893219, 'rewards/total_composite/std': 0.1720762699842453, 'reward': 0.10954989492893219, 'reward_std': 0.1720762699842453, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1407819241285324, 'sampling/sampling_logp_difference/max': 2.5419509410858154, 'sampling/importance_sampling_ratio/min': 0.07871268689632416, 'sampling/importance_sampling_ratio/mean': 1.023868441581726, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9626564495265484, 'clip_ratio/low_mean': 0.09288047160953283, 'clip_ratio/low_min': 0.09288047160953283, 'clip_ratio/high_mean': 0.0379081629216671, 'clip_ratio/high_max': 0.0379081629216671, 'clip_ratio/region_mean': 0.13078863453119993, 'reward_total_mean': 0.10954989492893219, 'reward_meter_mean': 0.27364253997802734, 'reward_meter_std': 0.4285583198070526, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7717296481132507, 'reward_repeat_penalty_std': 0.24070711433887482, 'reward_near_duplicate_penalty_mean': 0.9155372381210327, 'reward_near_duplicate_penalty_std': 0.057281166315078735, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9046062231063843, 'reward_distinct_2_std': 0.15360285341739655, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.10954989492893219, 'reward_total_composite_std': 0.1720762699842453, 'epoch': 0.03}

 24%|██▍       | 806/3300 [3:13:52<13:32:08, 19.54s/it]INFO 04-12 14:00:05 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 807/3300 [3:14:06<12:14:22, 17.67s/it]2026-04-12 14:00:18,176 | INFO | train_grpo_train | metrics_logged mode=train step=807

                                                       
{'loss': -0.1041, 'grad_norm': 2.359036445617676, 'learning_rate': 7.557575757575758e-06, 'num_tokens': 1639638.0, 'completions/mean_length': 170.125, 'completions/min_length': 51.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 56.16666793823242, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.4385441541671753, 'rewards/meter/std': 0.3289654552936554, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9390053153038025, 'rewards/lexical_plausibility/std': 0.08770473301410675, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.1505940705537796, 'rewards/repeat_penalty/mean': 0.7213979959487915, 'rewards/repeat_penalty/std': 0.09601594507694244, 'rewards/near_duplicate_penalty/mean': 0.9701355695724487, 'rewards/near_duplicate_penalty/std': 0.01984023116528988, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9508130550384521, 'rewards/distinct_2/std': 0.03979373723268509, 'rewards/total_composite/mean': 0.1577109694480896, 'rewards/total_composite/std': 0.11559614539146423, 'reward': 0.1577109694480896, 'reward_std': 0.11559613049030304, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10493512451648712, 'sampling/sampling_logp_difference/max': 1.499145746231079, 'sampling/importance_sampling_ratio/min': 0.223320871591568, 'sampling/importance_sampling_ratio/mean': 1.006955623626709, 'sampling/importance_sampling_ratio/max': 1.9603747129440308, 'entropy': 0.48986345902085304, 'clip_ratio/low_mean': 0.025925926864147186, 'clip_ratio/low_min': 0.025925926864147186, 'clip_ratio/high_mean': 0.049597340170294046, 'clip_ratio/high_max': 0.049597340170294046, 'clip_ratio/region_mean': 0.07552326703444123, 'reward_total_mean': 0.1577109694480896, 'reward_meter_mean': 0.4385441541671753, 'reward_meter_std': 0.3289654552936554, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9390053153038025, 'reward_lexical_plausibility_std': 0.08770473301410675, 'reward_repeat_penalty_mean': 0.7213979959487915, 'reward_repeat_penalty_std': 0.09601594507694244, 'reward_near_duplicate_penalty_mean': 0.9701355695724487, 'reward_near_duplicate_penalty_std': 0.01984023116528988, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9508130550384521, 'reward_distinct_2_std': 0.03979373723268509, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.1505940705537796, 'reward_total_composite_mean': 0.1577109694480896, 'reward_total_composite_std': 0.11559614539146423, 'epoch': 0.03}

 24%|██▍       | 807/3300 [3:14:06<12:14:22, 17.67s/it]INFO 04-12 14:00:18 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 808/3300 [3:14:19<11:20:26, 16.38s/it]2026-04-12 14:00:31,544 | INFO | train_grpo_train | metrics_logged mode=train step=808

                                                       
{'loss': 0.0102, 'grad_norm': 3.2151262760162354, 'learning_rate': 7.5545454545454555e-06, 'num_tokens': 1641044.0, 'completions/mean_length': 82.75, 'completions/min_length': 17.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 21.428571701049805, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.25902217626571655, 'rewards/meter/std': 0.4510461091995239, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9760870933532715, 'rewards/lexical_plausibility/std': 0.06763587146997452, 'rewards/judge_quality/mean': 0.4699999988079071, 'rewards/judge_quality/std': 0.3931557536125183, 'rewards/repeat_penalty/mean': 0.7457716464996338, 'rewards/repeat_penalty/std': 0.11551118642091751, 'rewards/near_duplicate_penalty/mean': 0.9526956081390381, 'rewards/near_duplicate_penalty/std': 0.07298299670219421, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.13849323987960815, 'rewards/total_composite/std': 0.31568512320518494, 'reward': 0.13849323987960815, 'reward_std': 0.31568509340286255, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18987688422203064, 'sampling/sampling_logp_difference/max': 4.385435104370117, 'sampling/importance_sampling_ratio/min': 0.012457466684281826, 'sampling/importance_sampling_ratio/mean': 1.0182334184646606, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6999455392360687, 'clip_ratio/low_mean': 0.07515787286683917, 'clip_ratio/low_min': 0.07515787286683917, 'clip_ratio/high_mean': 0.013194444589316845, 'clip_ratio/high_max': 0.013194444589316845, 'clip_ratio/region_mean': 0.08835231745615602, 'reward_total_mean': 0.13849323987960815, 'reward_meter_mean': 0.25902217626571655, 'reward_meter_std': 0.4510461091995239, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9760870933532715, 'reward_lexical_plausibility_std': 0.06763587146997452, 'reward_repeat_penalty_mean': 0.7457716464996338, 'reward_repeat_penalty_std': 0.11551118642091751, 'reward_near_duplicate_penalty_mean': 0.9526956081390381, 'reward_near_duplicate_penalty_std': 0.07298299670219421, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4699999988079071, 'reward_judge_quality_std': 0.3931557536125183, 'reward_total_composite_mean': 0.13849323987960815, 'reward_total_composite_std': 0.31568512320518494, 'epoch': 0.03}

 24%|██▍       | 808/3300 [3:14:19<11:20:26, 16.38s/it]INFO 04-12 14:00:31 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 809/3300 [3:14:29<10:04:37, 14.56s/it]2026-04-12 14:00:41,865 | INFO | train_grpo_train | metrics_logged mode=train step=809

                                                       
{'loss': 0.0209, 'grad_norm': 7.382968902587891, 'learning_rate': 7.551515151515152e-06, 'num_tokens': 1643827.0, 'completions/mean_length': 140.875, 'completions/min_length': 133.0, 'completions/max_length': 149.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 140.875, 'completions/min_terminated_length': 133.0, 'completions/max_terminated_length': 149.0, 'rewards/meter/mean': 0.8070170283317566, 'rewards/meter/std': 0.2273350954055786, '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/lexical_plausibility/mean': 0.9989224076271057, 'rewards/lexical_plausibility/std': 0.0030478707049041986, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.8352359533309937, 'rewards/repeat_penalty/std': 0.05035965517163277, 'rewards/near_duplicate_penalty/mean': 0.9464985728263855, 'rewards/near_duplicate_penalty/std': 0.024888422340154648, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8819787502288818, 'rewards/distinct_2/std': 0.03616122156381607, 'rewards/total_composite/mean': 0.2557677924633026, 'rewards/total_composite/std': 0.0903015211224556, 'reward': 0.2557677924633026, 'reward_std': 0.0903015211224556, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10854829847812653, 'sampling/sampling_logp_difference/max': 2.001067638397217, 'sampling/importance_sampling_ratio/min': 0.13519087433815002, 'sampling/importance_sampling_ratio/mean': 1.0040498971939087, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6414512395858765, 'clip_ratio/low_mean': 0.031189512461423874, 'clip_ratio/low_min': 0.031189512461423874, 'clip_ratio/high_mean': 0.07563150580972433, 'clip_ratio/high_max': 0.07563150580972433, 'clip_ratio/region_mean': 0.1068210182711482, 'reward_total_mean': 0.2557677924633026, 'reward_meter_mean': 0.8070170283317566, 'reward_meter_std': 0.2273350954055786, '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_lexical_plausibility_mean': 0.9989224076271057, 'reward_lexical_plausibility_std': 0.0030478707049041986, 'reward_repeat_penalty_mean': 0.8352359533309937, 'reward_repeat_penalty_std': 0.05035965517163277, 'reward_near_duplicate_penalty_mean': 0.9464985728263855, 'reward_near_duplicate_penalty_std': 0.024888422340154648, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8819787502288818, 'reward_distinct_2_std': 0.03616122156381607, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.2557677924633026, 'reward_total_composite_std': 0.0903015211224556, 'epoch': 0.03}

 25%|██▍       | 809/3300 [3:14:29<10:04:37, 14.56s/it]INFO 04-12 14:00:42 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 810/3300 [3:14:37<8:43:17, 12.61s/it] 2026-04-12 14:00:49,910 | INFO | train_grpo_train | metrics_logged mode=train step=810

                                                      
{'loss': 0.0607, 'grad_norm': 15.186004638671875, 'learning_rate': 7.548484848484849e-06, 'num_tokens': 1645482.0, 'completions/mean_length': 49.875, 'completions/min_length': 41.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.875, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.6428830623626709, 'rewards/meter/std': 0.3336312770843506, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8600000143051147, 'rewards/judge_quality/std': 0.2066052109003067, 'rewards/repeat_penalty/mean': 0.9832959771156311, 'rewards/repeat_penalty/std': 0.023197973147034645, 'rewards/near_duplicate_penalty/mean': 0.9832959771156311, 'rewards/near_duplicate_penalty/std': 0.023197973147034645, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5680773854255676, 'rewards/total_composite/std': 0.35162076354026794, 'reward': 0.5680773854255676, 'reward_std': 0.35162076354026794, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12980584800243378, 'sampling/sampling_logp_difference/max': 1.2337841987609863, 'sampling/importance_sampling_ratio/min': 0.2911885678768158, 'sampling/importance_sampling_ratio/mean': 1.0130921602249146, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7987408712506294, 'clip_ratio/low_mean': 0.0774453654885292, 'clip_ratio/low_min': 0.0774453654885292, 'clip_ratio/high_mean': 0.05865356419235468, 'clip_ratio/high_max': 0.05865356419235468, 'clip_ratio/region_mean': 0.13609892968088388, 'reward_total_mean': 0.5680773854255676, 'reward_meter_mean': 0.6428830623626709, 'reward_meter_std': 0.3336312770843506, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9832959771156311, 'reward_repeat_penalty_std': 0.023197973147034645, 'reward_near_duplicate_penalty_mean': 0.9832959771156311, 'reward_near_duplicate_penalty_std': 0.023197973147034645, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8600000143051147, 'reward_judge_quality_std': 0.2066052109003067, 'reward_total_composite_mean': 0.5680773854255676, 'reward_total_composite_std': 0.35162076354026794, 'epoch': 0.03}

 25%|██▍       | 810/3300 [3:14:37<8:43:17, 12.61s/it]INFO 04-12 14:00:50 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 811/3300 [3:14:48<8:22:38, 12.12s/it]2026-04-12 14:01:00,878 | INFO | train_grpo_train | metrics_logged mode=train step=811

                                                      
{'loss': 0.0752, 'grad_norm': 7.575557708740234, 'learning_rate': 7.545454545454546e-06, 'num_tokens': 1648527.0, 'completions/mean_length': 145.625, 'completions/min_length': 133.0, 'completions/max_length': 173.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 145.625, 'completions/min_terminated_length': 133.0, 'completions/max_terminated_length': 173.0, 'rewards/meter/mean': 0.4031369090080261, 'rewards/meter/std': 0.16512039303779602, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.8054443597793579, 'rewards/repeat_penalty/std': 0.20166487991809845, 'rewards/near_duplicate_penalty/mean': 0.9416308403015137, 'rewards/near_duplicate_penalty/std': 0.026360351592302322, 'rewards/opening_diversity/mean': 0.987500011920929, 'rewards/opening_diversity/std': 0.0353553481400013, 'rewards/distinct_2/mean': 0.8756292462348938, 'rewards/distinct_2/std': 0.15422005951404572, 'rewards/total_composite/mean': 0.13066896796226501, 'rewards/total_composite/std': 0.07025513797998428, 'reward': 0.13066896796226501, 'reward_std': 0.07025514543056488, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12699490785598755, 'sampling/sampling_logp_difference/max': 2.679246664047241, 'sampling/importance_sampling_ratio/min': 0.06861482560634613, 'sampling/importance_sampling_ratio/mean': 1.0148481130599976, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.666357297450304, 'clip_ratio/low_mean': 0.081703987903893, 'clip_ratio/low_min': 0.081703987903893, 'clip_ratio/high_mean': 0.056133151054382324, 'clip_ratio/high_max': 0.056133151054382324, 'clip_ratio/region_mean': 0.13783713895827532, 'reward_total_mean': 0.13066896796226501, 'reward_meter_mean': 0.4031369090080261, 'reward_meter_std': 0.16512039303779602, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8054443597793579, 'reward_repeat_penalty_std': 0.20166487991809845, 'reward_near_duplicate_penalty_mean': 0.9416308403015137, 'reward_near_duplicate_penalty_std': 0.026360351592302322, 'reward_opening_diversity_mean': 0.987500011920929, 'reward_opening_diversity_std': 0.0353553481400013, 'reward_distinct_2_mean': 0.8756292462348938, 'reward_distinct_2_std': 0.15422005951404572, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.13066896796226501, 'reward_total_composite_std': 0.07025513797998428, 'epoch': 0.03}

 25%|██▍       | 811/3300 [3:14:48<8:22:38, 12.12s/it]INFO 04-12 14:01:01 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 812/3300 [3:15:02<8:39:43, 12.53s/it]2026-04-12 14:01:14,387 | INFO | train_grpo_train | metrics_logged mode=train step=812

                                                      
{'loss': -0.0989, 'grad_norm': 3.800428867340088, 'learning_rate': 7.542424242424244e-06, 'num_tokens': 1650038.0, 'completions/mean_length': 106.875, 'completions/min_length': 47.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 49.000003814697266, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.7874910235404968, 'rewards/meter/std': 0.3190709948539734, '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/lexical_plausibility/mean': 0.9728875160217285, 'rewards/lexical_plausibility/std': 0.0766855999827385, 'rewards/judge_quality/mean': 0.6812499761581421, 'rewards/judge_quality/std': 0.34778639674186707, 'rewards/repeat_penalty/mean': 0.9951555728912354, 'rewards/repeat_penalty/std': 0.009139057248830795, 'rewards/near_duplicate_penalty/mean': 0.9951555728912354, 'rewards/near_duplicate_penalty/std': 0.009139057248830795, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6066259145736694, 'rewards/total_composite/std': 0.33748865127563477, 'reward': 0.6066259145736694, 'reward_std': 0.3374886214733124, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11039450764656067, 'sampling/sampling_logp_difference/max': 1.1034941673278809, 'sampling/importance_sampling_ratio/min': 0.33171001076698303, 'sampling/importance_sampling_ratio/mean': 1.0036505460739136, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5837102606892586, 'clip_ratio/low_mean': 0.03169326204806566, 'clip_ratio/low_min': 0.03169326204806566, 'clip_ratio/high_mean': 0.07745151408016682, 'clip_ratio/high_max': 0.07745151408016682, 'clip_ratio/region_mean': 0.10914477612823248, 'reward_total_mean': 0.6066259145736694, 'reward_meter_mean': 0.7874910235404968, 'reward_meter_std': 0.3190709948539734, '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_lexical_plausibility_mean': 0.9728875160217285, 'reward_lexical_plausibility_std': 0.0766855999827385, 'reward_repeat_penalty_mean': 0.9951555728912354, 'reward_repeat_penalty_std': 0.009139057248830795, 'reward_near_duplicate_penalty_mean': 0.9951555728912354, 'reward_near_duplicate_penalty_std': 0.009139057248830795, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6812499761581421, 'reward_judge_quality_std': 0.34778639674186707, 'reward_total_composite_mean': 0.6066259145736694, 'reward_total_composite_std': 0.33748865127563477, 'epoch': 0.03}

 25%|██▍       | 812/3300 [3:15:02<8:39:43, 12.53s/it]INFO 04-12 14:01:14 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 813/3300 [3:15:15<8:51:42, 12.83s/it]2026-04-12 14:01:27,902 | INFO | train_grpo_train | metrics_logged mode=train step=813

                                                      
{'loss': -0.1692, 'grad_norm': 3.3232412338256836, 'learning_rate': 7.53939393939394e-06, 'num_tokens': 1652949.0, 'completions/mean_length': 235.875, 'completions/min_length': 178.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 196.42857360839844, 'completions/min_terminated_length': 178.0, 'completions/max_terminated_length': 222.0, 'rewards/meter/mean': 0.9412060976028442, 'rewards/meter/std': 0.09580641239881516, 'rewards/count_adherence/mean': 0.78125, 'rewards/count_adherence/std': 0.2651650309562683, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9826085567474365, 'rewards/lexical_plausibility/std': 0.049190327525138855, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.4942803680896759, 'rewards/repeat_penalty/std': 0.28950735926628113, 'rewards/near_duplicate_penalty/mean': 0.8248645067214966, 'rewards/near_duplicate_penalty/std': 0.13155001401901245, 'rewards/opening_diversity/mean': 0.9508928656578064, 'rewards/opening_diversity/std': 0.08307480067014694, 'rewards/distinct_2/mean': 0.576888382434845, 'rewards/distinct_2/std': 0.27701079845428467, 'rewards/total_composite/mean': 0.21751300990581512, 'rewards/total_composite/std': 0.12358205765485764, 'reward': 0.21751300990581512, 'reward_std': 0.12358205020427704, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08466082066297531, 'sampling/sampling_logp_difference/max': 2.178527355194092, 'sampling/importance_sampling_ratio/min': 0.11320812255144119, 'sampling/importance_sampling_ratio/mean': 1.0056116580963135, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3652648665010929, 'clip_ratio/low_mean': 0.02692485461011529, 'clip_ratio/low_min': 0.02692485461011529, 'clip_ratio/high_mean': 0.03873121226206422, 'clip_ratio/high_max': 0.03873121226206422, 'clip_ratio/region_mean': 0.06565606687217951, 'reward_total_mean': 0.21751300990581512, 'reward_meter_mean': 0.9412060976028442, 'reward_meter_std': 0.09580641239881516, 'reward_count_adherence_mean': 0.78125, 'reward_count_adherence_std': 0.2651650309562683, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9826085567474365, 'reward_lexical_plausibility_std': 0.049190327525138855, 'reward_repeat_penalty_mean': 0.4942803680896759, 'reward_repeat_penalty_std': 0.28950735926628113, 'reward_near_duplicate_penalty_mean': 0.8248645067214966, 'reward_near_duplicate_penalty_std': 0.13155001401901245, 'reward_opening_diversity_mean': 0.9508928656578064, 'reward_opening_diversity_std': 0.08307480067014694, 'reward_distinct_2_mean': 0.576888382434845, 'reward_distinct_2_std': 0.27701079845428467, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.21751300990581512, 'reward_total_composite_std': 0.12358205765485764, 'epoch': 0.03}

 25%|██▍       | 813/3300 [3:15:15<8:51:42, 12.83s/it]INFO 04-12 14:01:28 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 814/3300 [3:15:29<8:56:03, 12.94s/it]2026-04-12 14:01:41,095 | INFO | train_grpo_train | metrics_logged mode=train step=814

                                                      
{'loss': -0.0491, 'grad_norm': 5.6867218017578125, 'learning_rate': 7.536363636363637e-06, 'num_tokens': 1655171.0, 'completions/mean_length': 156.75, 'completions/min_length': 94.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 106.00000762939453, 'completions/min_terminated_length': 94.0, 'completions/max_terminated_length': 128.0, 'rewards/meter/mean': 0.46127668023109436, 'rewards/meter/std': 0.31799080967903137, 'rewards/count_adherence/mean': 0.8500000238418579, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9457124471664429, 'rewards/lexical_plausibility/std': 0.12819664180278778, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.8547135591506958, 'rewards/repeat_penalty/std': 0.09622027724981308, 'rewards/near_duplicate_penalty/mean': 0.9376049041748047, 'rewards/near_duplicate_penalty/std': 0.03273974359035492, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.9166576862335205, 'rewards/distinct_2/std': 0.07167264819145203, 'rewards/total_composite/mean': 0.09750284254550934, 'rewards/total_composite/std': 0.0943702831864357, 'reward': 0.09750284254550934, 'reward_std': 0.0943702831864357, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1176077127456665, 'sampling/sampling_logp_difference/max': 1.5749835968017578, 'sampling/importance_sampling_ratio/min': 0.20701095461845398, 'sampling/importance_sampling_ratio/mean': 1.0081753730773926, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5940612480044365, 'clip_ratio/low_mean': 0.07036321051418781, 'clip_ratio/low_min': 0.07036321051418781, 'clip_ratio/high_mean': 0.04490788094699383, 'clip_ratio/high_max': 0.04490788094699383, 'clip_ratio/region_mean': 0.11527109146118164, 'reward_total_mean': 0.09750284254550934, 'reward_meter_mean': 0.46127668023109436, 'reward_meter_std': 0.31799080967903137, 'reward_count_adherence_mean': 0.8500000238418579, 'reward_count_adherence_std': 0.09258200973272324, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9457124471664429, 'reward_lexical_plausibility_std': 0.12819664180278778, 'reward_repeat_penalty_mean': 0.8547135591506958, 'reward_repeat_penalty_std': 0.09622027724981308, 'reward_near_duplicate_penalty_mean': 0.9376049041748047, 'reward_near_duplicate_penalty_std': 0.03273974359035492, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.9166576862335205, 'reward_distinct_2_std': 0.07167264819145203, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.09750284254550934, 'reward_total_composite_std': 0.0943702831864357, 'epoch': 0.03}

 25%|██▍       | 814/3300 [3:15:29<8:56:03, 12.94s/it]INFO 04-12 14:01:41 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 815/3300 [3:15:43<9:14:11, 13.38s/it]2026-04-12 14:01:55,511 | INFO | train_grpo_train | metrics_logged mode=train step=815

                                                      
{'loss': -0.0806, 'grad_norm': 1.8392436504364014, 'learning_rate': 7.533333333333334e-06, 'num_tokens': 1656727.0, 'completions/mean_length': 223.5, 'completions/min_length': 46.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 50.400001525878906, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.6055207252502441, 'rewards/meter/std': 0.45552077889442444, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.9034929275512695, 'rewards/lexical_plausibility/std': 0.15283240377902985, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.16903086006641388, 'rewards/repeat_penalty/mean': 0.9737361669540405, 'rewards/repeat_penalty/std': 0.05313893407583237, 'rewards/near_duplicate_penalty/mean': 0.9830358028411865, 'rewards/near_duplicate_penalty/std': 0.027686437591910362, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9898785352706909, 'rewards/distinct_2/std': 0.028627805411815643, 'rewards/total_composite/mean': 0.18413877487182617, 'rewards/total_composite/std': 0.1905413419008255, 'reward': 0.18413877487182617, 'reward_std': 0.1905413120985031, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1374284327030182, 'sampling/sampling_logp_difference/max': 1.6445746421813965, 'sampling/importance_sampling_ratio/min': 0.1930946707725525, 'sampling/importance_sampling_ratio/mean': 1.0112168788909912, 'sampling/importance_sampling_ratio/max': 1.7742395401000977, 'entropy': 0.5638574585318565, 'clip_ratio/low_mean': 0.0026041667442768812, 'clip_ratio/low_min': 0.0026041667442768812, 'clip_ratio/high_mean': 0.06918557081371546, 'clip_ratio/high_max': 0.06918557081371546, 'clip_ratio/region_mean': 0.07178973755799234, 'reward_total_mean': 0.18413877487182617, 'reward_meter_mean': 0.6055207252502441, 'reward_meter_std': 0.45552077889442444, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.9034929275512695, 'reward_lexical_plausibility_std': 0.15283240377902985, 'reward_repeat_penalty_mean': 0.9737361669540405, 'reward_repeat_penalty_std': 0.05313893407583237, 'reward_near_duplicate_penalty_mean': 0.9830358028411865, 'reward_near_duplicate_penalty_std': 0.027686437591910362, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9898785352706909, 'reward_distinct_2_std': 0.028627805411815643, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.16903086006641388, 'reward_total_composite_mean': 0.18413877487182617, 'reward_total_composite_std': 0.1905413419008255, 'epoch': 0.03}

 25%|██▍       | 815/3300 [3:15:43<9:14:11, 13.38s/it]INFO 04-12 14:01:55 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 816/3300 [3:15:51<8:03:59, 11.69s/it]2026-04-12 14:02:03,256 | INFO | train_grpo_train | metrics_logged mode=train step=816

                                                      
{'loss': 0.112, 'grad_norm': 24.046463012695312, 'learning_rate': 7.530303030303031e-06, 'num_tokens': 1658111.0, 'completions/mean_length': 24.0, 'completions/min_length': 21.0, 'completions/max_length': 31.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.0, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.8469799757003784, 'rewards/meter/std': 0.3430926501750946, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8525000214576721, 'rewards/judge_quality/std': 0.20331189036369324, 'rewards/repeat_penalty/mean': 0.7330512404441833, 'rewards/repeat_penalty/std': 0.032561879605054855, 'rewards/near_duplicate_penalty/mean': 0.977401614189148, 'rewards/near_duplicate_penalty/std': 0.043415818363428116, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.7126477360725403, 'rewards/total_composite/std': 0.34694704413414, 'reward': 0.7126477360725403, 'reward_std': 0.3469470143318176, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13056868314743042, 'sampling/sampling_logp_difference/max': 1.508354663848877, 'sampling/importance_sampling_ratio/min': 0.2212737500667572, 'sampling/importance_sampling_ratio/mean': 0.9936178922653198, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7504385411739349, 'clip_ratio/low_mean': 0.052769985049963, 'clip_ratio/low_min': 0.052769985049963, 'clip_ratio/high_mean': 0.11626353021711111, 'clip_ratio/high_max': 0.11626353021711111, 'clip_ratio/region_mean': 0.1690335152670741, 'reward_total_mean': 0.7126477360725403, 'reward_meter_mean': 0.8469799757003784, 'reward_meter_std': 0.3430926501750946, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7330512404441833, 'reward_repeat_penalty_std': 0.032561879605054855, 'reward_near_duplicate_penalty_mean': 0.977401614189148, 'reward_near_duplicate_penalty_std': 0.043415818363428116, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8525000214576721, 'reward_judge_quality_std': 0.20331189036369324, 'reward_total_composite_mean': 0.7126477360725403, 'reward_total_composite_std': 0.34694704413414, 'epoch': 0.03}

 25%|██▍       | 816/3300 [3:15:51<8:03:59, 11.69s/it]INFO 04-12 14:02:03 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 817/3300 [3:16:00<7:33:45, 10.96s/it]2026-04-12 14:02:12,526 | INFO | train_grpo_train | metrics_logged mode=train step=817

                                                      
{'loss': 0.0616, 'grad_norm': 9.506650924682617, 'learning_rate': 7.5272727272727274e-06, 'num_tokens': 1660075.0, 'completions/mean_length': 71.5, 'completions/min_length': 55.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.5, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.4503295421600342, 'rewards/meter/std': 0.37437984347343445, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5299999713897705, 'rewards/judge_quality/std': 0.24512389302253723, 'rewards/repeat_penalty/mean': 0.973711371421814, 'rewards/repeat_penalty/std': 0.030898524448275566, 'rewards/near_duplicate_penalty/mean': 0.9805337190628052, 'rewards/near_duplicate_penalty/std': 0.01512462180107832, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9928774833679199, 'rewards/distinct_2/std': 0.02014549821615219, 'rewards/total_composite/mean': 0.2287340760231018, 'rewards/total_composite/std': 0.207951158285141, 'reward': 0.2287340760231018, 'reward_std': 0.207951158285141, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11539392173290253, 'sampling/sampling_logp_difference/max': 1.0871162414550781, 'sampling/importance_sampling_ratio/min': 0.3371874690055847, 'sampling/importance_sampling_ratio/mean': 1.0085090398788452, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7391100414097309, 'clip_ratio/low_mean': 0.06265426147729158, 'clip_ratio/low_min': 0.06265426147729158, 'clip_ratio/high_mean': 0.02900556055828929, 'clip_ratio/high_max': 0.02900556055828929, 'clip_ratio/region_mean': 0.09165982203558087, 'reward_total_mean': 0.2287340760231018, 'reward_meter_mean': 0.4503295421600342, 'reward_meter_std': 0.37437984347343445, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.973711371421814, 'reward_repeat_penalty_std': 0.030898524448275566, 'reward_near_duplicate_penalty_mean': 0.9805337190628052, 'reward_near_duplicate_penalty_std': 0.01512462180107832, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9928774833679199, 'reward_distinct_2_std': 0.02014549821615219, 'reward_judge_quality_mean': 0.5299999713897705, 'reward_judge_quality_std': 0.24512389302253723, 'reward_total_composite_mean': 0.2287340760231018, 'reward_total_composite_std': 0.207951158285141, 'epoch': 0.03}

 25%|██▍       | 817/3300 [3:16:00<7:33:45, 10.96s/it]INFO 04-12 14:02:12 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 818/3300 [3:16:11<7:37:22, 11.06s/it]2026-04-12 14:02:23,797 | INFO | train_grpo_train | metrics_logged mode=train step=818

                                                      
{'loss': 0.0361, 'grad_norm': 4.781548976898193, 'learning_rate': 7.524242424242425e-06, 'num_tokens': 1663616.0, 'completions/mean_length': 252.625, 'completions/min_length': 218.0, 'completions/max_length': 288.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 252.625, 'completions/min_terminated_length': 218.0, 'completions/max_terminated_length': 288.0, 'rewards/meter/mean': 0.8705605268478394, 'rewards/meter/std': 0.20197537541389465, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.24633942544460297, 'rewards/repeat_penalty/std': 0.16943033039569855, 'rewards/near_duplicate_penalty/mean': 0.7904670238494873, 'rewards/near_duplicate_penalty/std': 0.12964604794979095, 'rewards/opening_diversity/mean': 0.9895833730697632, 'rewards/opening_diversity/std': 0.029462775215506554, 'rewards/distinct_2/mean': 0.32842856645584106, 'rewards/distinct_2/std': 0.2051476538181305, 'rewards/total_composite/mean': 0.23434017598628998, 'rewards/total_composite/std': 0.10898102074861526, 'reward': 0.23434017598628998, 'reward_std': 0.10898101329803467, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07445370405912399, 'sampling/sampling_logp_difference/max': 2.2147040367126465, 'sampling/importance_sampling_ratio/min': 0.10918582230806351, 'sampling/importance_sampling_ratio/mean': 0.998491108417511, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.37534728460013866, 'clip_ratio/low_mean': 0.03056690702214837, 'clip_ratio/low_min': 0.03056690702214837, 'clip_ratio/high_mean': 0.04160299897193909, 'clip_ratio/high_max': 0.04160299897193909, 'clip_ratio/region_mean': 0.07216990599408746, 'reward_total_mean': 0.23434017598628998, 'reward_meter_mean': 0.8705605268478394, 'reward_meter_std': 0.20197537541389465, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.24633942544460297, 'reward_repeat_penalty_std': 0.16943033039569855, 'reward_near_duplicate_penalty_mean': 0.7904670238494873, 'reward_near_duplicate_penalty_std': 0.12964604794979095, 'reward_opening_diversity_mean': 0.9895833730697632, 'reward_opening_diversity_std': 0.029462775215506554, 'reward_distinct_2_mean': 0.32842856645584106, 'reward_distinct_2_std': 0.2051476538181305, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.23434017598628998, 'reward_total_composite_std': 0.10898102074861526, 'epoch': 0.03}

 25%|██▍       | 818/3300 [3:16:11<7:37:22, 11.06s/it]INFO 04-12 14:02:24 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 819/3300 [3:16:25<8:09:59, 11.85s/it]2026-04-12 14:02:37,504 | INFO | train_grpo_train | metrics_logged mode=train step=819

                                                      
{'loss': -0.0568, 'grad_norm': 2.4939849376678467, 'learning_rate': 7.521212121212121e-06, 'num_tokens': 1665105.0, 'completions/mean_length': 222.125, 'completions/min_length': 43.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 48.20000076293945, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.48654890060424805, 'rewards/meter/std': 0.4777888357639313, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.9322326183319092, 'rewards/lexical_plausibility/std': 0.09612184762954712, 'rewards/judge_quality/mean': 0.4125000238418579, 'rewards/judge_quality/std': 0.35346245765686035, 'rewards/repeat_penalty/mean': 0.998511016368866, 'rewards/repeat_penalty/std': 0.0029533349443227053, 'rewards/near_duplicate_penalty/mean': 0.9990079402923584, 'rewards/near_duplicate_penalty/std': 0.0028059897013008595, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9995030760765076, 'rewards/distinct_2/std': 0.0014054919593036175, 'rewards/total_composite/mean': 0.21199412643909454, 'rewards/total_composite/std': 0.28878629207611084, 'reward': 0.21199412643909454, 'reward_std': 0.28878626227378845, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14067530632019043, 'sampling/sampling_logp_difference/max': 1.3530428409576416, 'sampling/importance_sampling_ratio/min': 0.2584526240825653, 'sampling/importance_sampling_ratio/mean': 0.991131603717804, 'sampling/importance_sampling_ratio/max': 1.9702669382095337, 'entropy': 0.5373192317783833, 'clip_ratio/low_mean': 0.01886792480945587, 'clip_ratio/low_min': 0.01886792480945587, 'clip_ratio/high_mean': 0.0463135102763772, 'clip_ratio/high_max': 0.0463135102763772, 'clip_ratio/region_mean': 0.06518143508583307, 'reward_total_mean': 0.21199412643909454, 'reward_meter_mean': 0.48654890060424805, 'reward_meter_std': 0.4777888357639313, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.9322326183319092, 'reward_lexical_plausibility_std': 0.09612184762954712, 'reward_repeat_penalty_mean': 0.998511016368866, 'reward_repeat_penalty_std': 0.0029533349443227053, 'reward_near_duplicate_penalty_mean': 0.9990079402923584, 'reward_near_duplicate_penalty_std': 0.0028059897013008595, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9995030760765076, 'reward_distinct_2_std': 0.0014054919593036175, 'reward_judge_quality_mean': 0.4125000238418579, 'reward_judge_quality_std': 0.35346245765686035, 'reward_total_composite_mean': 0.21199412643909454, 'reward_total_composite_std': 0.28878629207611084, 'epoch': 0.03}

 25%|██▍       | 819/3300 [3:16:25<8:09:59, 11.85s/it]INFO 04-12 14:02:37 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 820/3300 [3:16:34<7:38:38, 11.10s/it]2026-04-12 14:02:46,839 | INFO | train_grpo_train | metrics_logged mode=train step=820

                                                      
{'loss': 0.0655, 'grad_norm': 8.09565258026123, 'learning_rate': 7.518181818181819e-06, 'num_tokens': 1667433.0, 'completions/mean_length': 113.0, 'completions/min_length': 97.0, 'completions/max_length': 123.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 113.0, 'completions/min_terminated_length': 97.0, 'completions/max_terminated_length': 123.0, 'rewards/meter/mean': 0.5458306670188904, 'rewards/meter/std': 0.32007092237472534, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.44999998807907104, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.8199503421783447, 'rewards/repeat_penalty/std': 0.09975920617580414, 'rewards/near_duplicate_penalty/mean': 0.9429879188537598, 'rewards/near_duplicate_penalty/std': 0.03316446766257286, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.867276132106781, 'rewards/distinct_2/std': 0.07919946312904358, 'rewards/total_composite/mean': 0.18988703191280365, 'rewards/total_composite/std': 0.12961415946483612, 'reward': 0.18988703191280365, 'reward_std': 0.12961415946483612, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1332654356956482, 'sampling/sampling_logp_difference/max': 1.441481113433838, 'sampling/importance_sampling_ratio/min': 0.236577108502388, 'sampling/importance_sampling_ratio/mean': 1.0116586685180664, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7692822888493538, 'clip_ratio/low_mean': 0.05453959107398987, 'clip_ratio/low_min': 0.05453959107398987, 'clip_ratio/high_mean': 0.08198838122189045, 'clip_ratio/high_max': 0.08198838122189045, 'clip_ratio/region_mean': 0.13652797229588032, 'reward_total_mean': 0.18988703191280365, 'reward_meter_mean': 0.5458306670188904, 'reward_meter_std': 0.32007092237472534, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8199503421783447, 'reward_repeat_penalty_std': 0.09975920617580414, 'reward_near_duplicate_penalty_mean': 0.9429879188537598, 'reward_near_duplicate_penalty_std': 0.03316446766257286, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.867276132106781, 'reward_distinct_2_std': 0.07919946312904358, 'reward_judge_quality_mean': 0.44999998807907104, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.18988703191280365, 'reward_total_composite_std': 0.12961415946483612, 'epoch': 0.03}

 25%|██▍       | 820/3300 [3:16:34<7:38:38, 11.10s/it]INFO 04-12 14:02:47 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 821/3300 [3:16:48<8:07:31, 11.80s/it]2026-04-12 14:03:00,279 | INFO | train_grpo_train | metrics_logged mode=train step=821

                                                      
{'loss': -0.0938, 'grad_norm': 2.969666004180908, 'learning_rate': 7.515151515151516e-06, 'num_tokens': 1669106.0, 'completions/mean_length': 168.125, 'completions/min_length': 46.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 53.5, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.6630713939666748, 'rewards/meter/std': 0.4266056418418884, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9276131391525269, 'rewards/lexical_plausibility/std': 0.13415464758872986, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.1767766922712326, 'rewards/repeat_penalty/mean': 0.9203839898109436, 'rewards/repeat_penalty/std': 0.10864054411649704, 'rewards/near_duplicate_penalty/mean': 0.9828839898109436, 'rewards/near_duplicate_penalty/std': 0.02924136444926262, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2532910704612732, 'rewards/total_composite/std': 0.19527983665466309, 'reward': 0.2532910704612732, 'reward_std': 0.19527983665466309, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1490723043680191, 'sampling/sampling_logp_difference/max': 1.6246337890625, 'sampling/importance_sampling_ratio/min': 0.19698379933834076, 'sampling/importance_sampling_ratio/mean': 1.0188796520233154, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6903570517897606, 'clip_ratio/low_mean': 0.007352941203862429, 'clip_ratio/low_min': 0.007352941203862429, 'clip_ratio/high_mean': 0.10040908213704824, 'clip_ratio/high_max': 0.10040908213704824, 'clip_ratio/region_mean': 0.10776202334091067, 'reward_total_mean': 0.2532910704612732, 'reward_meter_mean': 0.6630713939666748, 'reward_meter_std': 0.4266056418418884, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9276131391525269, 'reward_lexical_plausibility_std': 0.13415464758872986, 'reward_repeat_penalty_mean': 0.9203839898109436, 'reward_repeat_penalty_std': 0.10864054411649704, 'reward_near_duplicate_penalty_mean': 0.9828839898109436, 'reward_near_duplicate_penalty_std': 0.02924136444926262, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.1767766922712326, 'reward_total_composite_mean': 0.2532910704612732, 'reward_total_composite_std': 0.19527983665466309, 'epoch': 0.03}

 25%|██▍       | 821/3300 [3:16:48<8:07:31, 11.80s/it]INFO 04-12 14:03:00 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 822/3300 [3:17:01<8:26:43, 12.27s/it]2026-04-12 14:03:13,642 | INFO | train_grpo_train | metrics_logged mode=train step=822

                                                      
{'loss': -0.0417, 'grad_norm': 4.708151817321777, 'learning_rate': 7.512121212121213e-06, 'num_tokens': 1670805.0, 'completions/mean_length': 103.375, 'completions/min_length': 43.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 45.000003814697266, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.3182946443557739, 'rewards/meter/std': 0.3539869785308838, '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/lexical_plausibility/mean': 0.9741119146347046, 'rewards/lexical_plausibility/std': 0.07322254031896591, 'rewards/judge_quality/mean': 0.752500057220459, 'rewards/judge_quality/std': 0.32805708050727844, 'rewards/repeat_penalty/mean': 0.9276940226554871, 'rewards/repeat_penalty/std': 0.09563203155994415, 'rewards/near_duplicate_penalty/mean': 0.9678204655647278, 'rewards/near_duplicate_penalty/std': 0.039478566497564316, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.956501841545105, 'rewards/distinct_2/std': 0.0633978545665741, 'rewards/total_composite/mean': 0.28315919637680054, 'rewards/total_composite/std': 0.33432263135910034, 'reward': 0.28315919637680054, 'reward_std': 0.33432263135910034, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12824253737926483, 'sampling/sampling_logp_difference/max': 1.1638872623443604, 'sampling/importance_sampling_ratio/min': 0.312269926071167, 'sampling/importance_sampling_ratio/mean': 1.0155647993087769, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5962987914681435, 'clip_ratio/low_mean': 0.061575678177177906, 'clip_ratio/low_min': 0.061575678177177906, 'clip_ratio/high_mean': 0.05713362991809845, 'clip_ratio/high_max': 0.05713362991809845, 'clip_ratio/region_mean': 0.11870930809527636, 'reward_total_mean': 0.28315919637680054, 'reward_meter_mean': 0.3182946443557739, 'reward_meter_std': 0.3539869785308838, '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_lexical_plausibility_mean': 0.9741119146347046, 'reward_lexical_plausibility_std': 0.07322254031896591, 'reward_repeat_penalty_mean': 0.9276940226554871, 'reward_repeat_penalty_std': 0.09563203155994415, 'reward_near_duplicate_penalty_mean': 0.9678204655647278, 'reward_near_duplicate_penalty_std': 0.039478566497564316, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.956501841545105, 'reward_distinct_2_std': 0.0633978545665741, 'reward_judge_quality_mean': 0.752500057220459, 'reward_judge_quality_std': 0.32805708050727844, 'reward_total_composite_mean': 0.28315919637680054, 'reward_total_composite_std': 0.33432263135910034, 'epoch': 0.03}

 25%|██▍       | 822/3300 [3:17:01<8:26:43, 12.27s/it]INFO 04-12 14:03:13 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 823/3300 [3:17:15<8:51:48, 12.88s/it]2026-04-12 14:03:27,959 | INFO | train_grpo_train | metrics_logged mode=train step=823

                                                      
{'loss': -0.1665, 'grad_norm': 3.7394113540649414, 'learning_rate': 7.509090909090909e-06, 'num_tokens': 1673552.0, 'completions/mean_length': 230.375, 'completions/min_length': 167.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 190.1428680419922, 'completions/min_terminated_length': 167.0, 'completions/max_terminated_length': 232.0, 'rewards/meter/mean': 0.9197593927383423, 'rewards/meter/std': 0.1841699331998825, 'rewards/count_adherence/mean': 0.8035714626312256, 'rewards/count_adherence/std': 0.215134397149086, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9680410027503967, 'rewards/lexical_plausibility/std': 0.09039372205734253, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.11649647355079651, 'rewards/repeat_penalty/mean': 0.42832398414611816, 'rewards/repeat_penalty/std': 0.24702930450439453, 'rewards/near_duplicate_penalty/mean': 0.7599433660507202, 'rewards/near_duplicate_penalty/std': 0.24937529861927032, 'rewards/opening_diversity/mean': 0.859375, 'rewards/opening_diversity/std': 0.30935922265052795, 'rewards/distinct_2/mean': 0.5753481984138489, 'rewards/distinct_2/std': 0.3082761764526367, 'rewards/total_composite/mean': 0.17166483402252197, 'rewards/total_composite/std': 0.09905346482992172, 'reward': 0.17166483402252197, 'reward_std': 0.09905345737934113, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10652532428503036, 'sampling/sampling_logp_difference/max': 2.2755303382873535, 'sampling/importance_sampling_ratio/min': 0.10274240374565125, 'sampling/importance_sampling_ratio/mean': 1.0041534900665283, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5771142337471247, 'clip_ratio/low_mean': 0.032494738698005676, 'clip_ratio/low_min': 0.032494738698005676, 'clip_ratio/high_mean': 0.04119925294071436, 'clip_ratio/high_max': 0.04119925294071436, 'clip_ratio/region_mean': 0.07369399163872004, 'reward_total_mean': 0.17166483402252197, 'reward_meter_mean': 0.9197593927383423, 'reward_meter_std': 0.1841699331998825, 'reward_count_adherence_mean': 0.8035714626312256, 'reward_count_adherence_std': 0.215134397149086, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9680410027503967, 'reward_lexical_plausibility_std': 0.09039372205734253, 'reward_repeat_penalty_mean': 0.42832398414611816, 'reward_repeat_penalty_std': 0.24702930450439453, 'reward_near_duplicate_penalty_mean': 0.7599433660507202, 'reward_near_duplicate_penalty_std': 0.24937529861927032, 'reward_opening_diversity_mean': 0.859375, 'reward_opening_diversity_std': 0.30935922265052795, 'reward_distinct_2_mean': 0.5753481984138489, 'reward_distinct_2_std': 0.3082761764526367, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.11649647355079651, 'reward_total_composite_mean': 0.17166483402252197, 'reward_total_composite_std': 0.09905346482992172, 'epoch': 0.03}

 25%|██▍       | 823/3300 [3:17:15<8:51:48, 12.88s/it]INFO 04-12 14:03:28 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 824/3300 [3:17:23<7:50:03, 11.39s/it]2026-04-12 14:03:35,872 | INFO | train_grpo_train | metrics_logged mode=train step=824

                                                      
{'loss': 0.0071, 'grad_norm': 17.515064239501953, 'learning_rate': 7.5060606060606065e-06, 'num_tokens': 1675218.0, 'completions/mean_length': 45.25, 'completions/min_length': 41.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.25, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.4593569040298462, 'rewards/meter/std': 0.4075322449207306, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6262500286102295, 'rewards/judge_quality/std': 0.2432481348514557, 'rewards/repeat_penalty/mean': 0.906800389289856, 'rewards/repeat_penalty/std': 0.1017073392868042, 'rewards/near_duplicate_penalty/mean': 0.9504185914993286, 'rewards/near_duplicate_penalty/std': 0.0443868413567543, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9518266916275024, 'rewards/distinct_2/std': 0.07075672596693039, 'rewards/total_composite/mean': 0.2819594740867615, 'rewards/total_composite/std': 0.27868616580963135, 'reward': 0.2819594740867615, 'reward_std': 0.27868616580963135, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15847836434841156, 'sampling/sampling_logp_difference/max': 1.7431327104568481, 'sampling/importance_sampling_ratio/min': 0.17497140169143677, 'sampling/importance_sampling_ratio/mean': 1.023474097251892, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9544377326965332, 'clip_ratio/low_mean': 0.08723642397671938, 'clip_ratio/low_min': 0.08723642397671938, 'clip_ratio/high_mean': 0.054752906784415245, 'clip_ratio/high_max': 0.054752906784415245, 'clip_ratio/region_mean': 0.14198933076113462, 'reward_total_mean': 0.2819594740867615, 'reward_meter_mean': 0.4593569040298462, 'reward_meter_std': 0.4075322449207306, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.906800389289856, 'reward_repeat_penalty_std': 0.1017073392868042, 'reward_near_duplicate_penalty_mean': 0.9504185914993286, 'reward_near_duplicate_penalty_std': 0.0443868413567543, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9518266916275024, 'reward_distinct_2_std': 0.07075672596693039, 'reward_judge_quality_mean': 0.6262500286102295, 'reward_judge_quality_std': 0.2432481348514557, 'reward_total_composite_mean': 0.2819594740867615, 'reward_total_composite_std': 0.27868616580963135, 'epoch': 0.03}

 25%|██▍       | 824/3300 [3:17:23<7:50:03, 11.39s/it]INFO 04-12 14:03:36 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 825/3300 [3:17:36<8:09:11, 11.86s/it]2026-04-12 14:03:48,820 | INFO | train_grpo_train | metrics_logged mode=train step=825

                                                      
{'loss': -0.0429, 'grad_norm': 2.90382719039917, 'learning_rate': 7.503030303030303e-06, 'num_tokens': 1676872.0, 'completions/mean_length': 166.75, 'completions/min_length': 50.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 51.66666793823242, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.504851758480072, 'rewards/meter/std': 0.4356744885444641, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9281660318374634, 'rewards/lexical_plausibility/std': 0.09923125803470612, 'rewards/judge_quality/mean': 0.2837499976158142, 'rewards/judge_quality/std': 0.28278905153274536, 'rewards/repeat_penalty/mean': 0.6415095329284668, 'rewards/repeat_penalty/std': 0.2112969309091568, 'rewards/near_duplicate_penalty/mean': 0.9031348824501038, 'rewards/near_duplicate_penalty/std': 0.05762649327516556, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.8525846600532532, 'rewards/distinct_2/std': 0.11896373331546783, 'rewards/total_composite/mean': 0.21184338629245758, 'rewards/total_composite/std': 0.3044293224811554, 'reward': 0.21184338629245758, 'reward_std': 0.3044293224811554, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12933504581451416, 'sampling/sampling_logp_difference/max': 1.1533470153808594, 'sampling/importance_sampling_ratio/min': 0.31557875871658325, 'sampling/importance_sampling_ratio/mean': 1.0194796323776245, 'sampling/importance_sampling_ratio/max': 1.8992120027542114, 'entropy': 0.8100524507462978, 'clip_ratio/low_mean': 0.054800222627818584, 'clip_ratio/low_min': 0.054800222627818584, 'clip_ratio/high_mean': 0.017500000074505806, 'clip_ratio/high_max': 0.017500000074505806, 'clip_ratio/region_mean': 0.07230022270232439, 'reward_total_mean': 0.21184338629245758, 'reward_meter_mean': 0.504851758480072, 'reward_meter_std': 0.4356744885444641, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9281660318374634, 'reward_lexical_plausibility_std': 0.09923125803470612, 'reward_repeat_penalty_mean': 0.6415095329284668, 'reward_repeat_penalty_std': 0.2112969309091568, 'reward_near_duplicate_penalty_mean': 0.9031348824501038, 'reward_near_duplicate_penalty_std': 0.05762649327516556, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.8525846600532532, 'reward_distinct_2_std': 0.11896373331546783, 'reward_judge_quality_mean': 0.2837499976158142, 'reward_judge_quality_std': 0.28278905153274536, 'reward_total_composite_mean': 0.21184338629245758, 'reward_total_composite_std': 0.3044293224811554, 'epoch': 0.03}

 25%|██▌       | 825/3300 [3:17:36<8:09:11, 11.86s/it]INFO 04-12 14:03:49 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 826/3300 [3:17:49<8:15:01, 12.01s/it]2026-04-12 14:04:01,167 | INFO | train_grpo_train | metrics_logged mode=train step=826

                                                      
{'loss': 0.093, 'grad_norm': 5.559487342834473, 'learning_rate': 7.500000000000001e-06, 'num_tokens': 1681173.0, 'completions/mean_length': 297.625, 'completions/min_length': 237.0, 'completions/max_length': 351.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 297.625, 'completions/min_terminated_length': 237.0, 'completions/max_terminated_length': 351.0, 'rewards/meter/mean': 0.8171510696411133, 'rewards/meter/std': 0.1934913843870163, 'rewards/count_adherence/mean': 0.8863636255264282, 'rewards/count_adherence/std': 0.06428244709968567, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.22500000894069672, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.3406097888946533, 'rewards/repeat_penalty/std': 0.31617647409439087, 'rewards/near_duplicate_penalty/mean': 0.8323519229888916, 'rewards/near_duplicate_penalty/std': 0.08623382449150085, 'rewards/opening_diversity/mean': 0.8984375, 'rewards/opening_diversity/std': 0.15434837341308594, 'rewards/distinct_2/mean': 0.42978453636169434, 'rewards/distinct_2/std': 0.3113824129104614, 'rewards/total_composite/mean': 0.16477039456367493, 'rewards/total_composite/std': 0.08743821084499359, 'reward': 0.16477039456367493, 'reward_std': 0.08743821829557419, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0844714343547821, 'sampling/sampling_logp_difference/max': 2.0492701530456543, 'sampling/importance_sampling_ratio/min': 0.12882889807224274, 'sampling/importance_sampling_ratio/mean': 1.010542869567871, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.46387164667248726, 'clip_ratio/low_mean': 0.040871933568269014, 'clip_ratio/low_min': 0.040871933568269014, 'clip_ratio/high_mean': 0.03977311495691538, 'clip_ratio/high_max': 0.03977311495691538, 'clip_ratio/region_mean': 0.08064504852518439, 'reward_total_mean': 0.16477039456367493, 'reward_meter_mean': 0.8171510696411133, 'reward_meter_std': 0.1934913843870163, 'reward_count_adherence_mean': 0.8863636255264282, 'reward_count_adherence_std': 0.06428244709968567, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.3406097888946533, 'reward_repeat_penalty_std': 0.31617647409439087, 'reward_near_duplicate_penalty_mean': 0.8323519229888916, 'reward_near_duplicate_penalty_std': 0.08623382449150085, 'reward_opening_diversity_mean': 0.8984375, 'reward_opening_diversity_std': 0.15434837341308594, 'reward_distinct_2_mean': 0.42978453636169434, 'reward_distinct_2_std': 0.3113824129104614, 'reward_judge_quality_mean': 0.22500000894069672, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.16477039456367493, 'reward_total_composite_std': 0.08743821084499359, 'epoch': 0.03}

 25%|██▌       | 826/3300 [3:17:49<8:15:01, 12.01s/it]INFO 04-12 14:04:01 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 827/3300 [3:17:57<7:28:19, 10.88s/it]2026-04-12 14:04:09,413 | INFO | train_grpo_train | metrics_logged mode=train step=827

                                                      
{'loss': 0.0612, 'grad_norm': 11.485149383544922, 'learning_rate': 7.496969696969698e-06, 'num_tokens': 1683029.0, 'completions/mean_length': 46.0, 'completions/min_length': 43.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.0, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.5837651491165161, 'rewards/meter/std': 0.363710880279541, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.8799999952316284, 'rewards/judge_quality/std': 0.1742740124464035, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5367031693458557, 'rewards/total_composite/std': 0.35172075033187866, 'reward': 0.5367031693458557, 'reward_std': 0.35172075033187866, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.108575239777565, 'sampling/sampling_logp_difference/max': 2.805960178375244, 'sampling/importance_sampling_ratio/min': 0.06044870242476463, 'sampling/importance_sampling_ratio/mean': 0.9819539189338684, 'sampling/importance_sampling_ratio/max': 1.9269944429397583, 'entropy': 0.330756351351738, 'clip_ratio/low_mean': 0.029451093869283795, 'clip_ratio/low_min': 0.029451093869283795, 'clip_ratio/high_mean': 0.050069005228579044, 'clip_ratio/high_max': 0.050069005228579044, 'clip_ratio/region_mean': 0.07952009909786284, 'reward_total_mean': 0.5367031693458557, 'reward_meter_mean': 0.5837651491165161, 'reward_meter_std': 0.363710880279541, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8799999952316284, 'reward_judge_quality_std': 0.1742740124464035, 'reward_total_composite_mean': 0.5367031693458557, 'reward_total_composite_std': 0.35172075033187866, 'epoch': 0.03}

 25%|██▌       | 827/3300 [3:17:57<7:28:19, 10.88s/it]INFO 04-12 14:04:09 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 828/3300 [3:18:05<6:52:03, 10.00s/it]2026-04-12 14:04:17,371 | INFO | train_grpo_train | metrics_logged mode=train step=828

                                                      
{'loss': 0.0348, 'grad_norm': 12.208457946777344, 'learning_rate': 7.493939393939395e-06, 'num_tokens': 1684714.0, 'completions/mean_length': 51.625, 'completions/min_length': 47.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.625, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.7387732863426208, 'rewards/meter/std': 0.3698543310165405, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6875, 'rewards/judge_quality/std': 0.24211864173412323, 'rewards/repeat_penalty/mean': 0.9855691194534302, 'rewards/repeat_penalty/std': 0.018926572054624557, 'rewards/near_duplicate_penalty/mean': 0.9855691194534302, 'rewards/near_duplicate_penalty/std': 0.018926572054624557, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4585193693637848, 'rewards/total_composite/std': 0.2535739839076996, 'reward': 0.4585193693637848, 'reward_std': 0.2535739839076996, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1267327219247818, 'sampling/sampling_logp_difference/max': 1.2790570259094238, 'sampling/importance_sampling_ratio/min': 0.27829962968826294, 'sampling/importance_sampling_ratio/mean': 1.0119009017944336, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8061745166778564, 'clip_ratio/low_mean': 0.07376865297555923, 'clip_ratio/low_min': 0.07376865297555923, 'clip_ratio/high_mean': 0.044817928690463305, 'clip_ratio/high_max': 0.044817928690463305, 'clip_ratio/region_mean': 0.11858658166602254, 'reward_total_mean': 0.4585193693637848, 'reward_meter_mean': 0.7387732863426208, 'reward_meter_std': 0.3698543310165405, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9855691194534302, 'reward_repeat_penalty_std': 0.018926572054624557, 'reward_near_duplicate_penalty_mean': 0.9855691194534302, 'reward_near_duplicate_penalty_std': 0.018926572054624557, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6875, 'reward_judge_quality_std': 0.24211864173412323, 'reward_total_composite_mean': 0.4585193693637848, 'reward_total_composite_std': 0.2535739839076996, 'epoch': 0.03}

 25%|██▌       | 828/3300 [3:18:05<6:52:03, 10.00s/it]INFO 04-12 14:04:17 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 829/3300 [3:18:18<7:35:37, 11.06s/it]2026-04-12 14:04:30,913 | INFO | train_grpo_train | metrics_logged mode=train step=829

                                                      
{'loss': -0.0491, 'grad_norm': 4.997721195220947, 'learning_rate': 7.490909090909092e-06, 'num_tokens': 1686790.0, 'completions/mean_length': 159.5, 'completions/min_length': 92.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 109.14286041259766, 'completions/min_terminated_length': 92.0, 'completions/max_terminated_length': 132.0, 'rewards/meter/mean': 0.7552764415740967, 'rewards/meter/std': 0.32096803188323975, '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/lexical_plausibility/mean': 0.9618983864784241, 'rewards/lexical_plausibility/std': 0.10776761174201965, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.5502966642379761, 'rewards/repeat_penalty/std': 0.30307191610336304, 'rewards/near_duplicate_penalty/mean': 0.8795453310012817, 'rewards/near_duplicate_penalty/std': 0.0521891750395298, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.6573268175125122, 'rewards/distinct_2/std': 0.3117270767688751, 'rewards/total_composite/mean': 0.21154771745204926, 'rewards/total_composite/std': 0.13740533590316772, 'reward': 0.21154771745204926, 'reward_std': 0.13740533590316772, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09593869745731354, 'sampling/sampling_logp_difference/max': 1.836510181427002, 'sampling/importance_sampling_ratio/min': 0.15937264263629913, 'sampling/importance_sampling_ratio/mean': 1.0106825828552246, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5000378489494324, 'clip_ratio/low_mean': 0.036383810453116894, 'clip_ratio/low_min': 0.036383810453116894, 'clip_ratio/high_mean': 0.05148236081004143, 'clip_ratio/high_max': 0.05148236081004143, 'clip_ratio/region_mean': 0.08786617126315832, 'reward_total_mean': 0.21154771745204926, 'reward_meter_mean': 0.7552764415740967, 'reward_meter_std': 0.32096803188323975, '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_lexical_plausibility_mean': 0.9618983864784241, 'reward_lexical_plausibility_std': 0.10776761174201965, 'reward_repeat_penalty_mean': 0.5502966642379761, 'reward_repeat_penalty_std': 0.30307191610336304, 'reward_near_duplicate_penalty_mean': 0.8795453310012817, 'reward_near_duplicate_penalty_std': 0.0521891750395298, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.6573268175125122, 'reward_distinct_2_std': 0.3117270767688751, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.21154771745204926, 'reward_total_composite_std': 0.13740533590316772, 'epoch': 0.03}

 25%|██▌       | 829/3300 [3:18:18<7:35:37, 11.06s/it]INFO 04-12 14:04:31 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 830/3300 [3:18:31<8:00:21, 11.67s/it]2026-04-12 14:04:43,990 | INFO | train_grpo_train | metrics_logged mode=train step=830

                                                      
{'loss': -0.0401, 'grad_norm': 3.4121365547180176, 'learning_rate': 7.487878787878788e-06, 'num_tokens': 1688406.0, 'completions/mean_length': 96.0, 'completions/min_length': 32.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 36.57143020629883, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.6927865743637085, 'rewards/meter/std': 0.3873361349105835, '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/lexical_plausibility/mean': 0.9700043797492981, 'rewards/lexical_plausibility/std': 0.08484045416116714, 'rewards/judge_quality/mean': 0.7637499570846558, 'rewards/judge_quality/std': 0.3349173665046692, 'rewards/repeat_penalty/mean': 0.998106062412262, 'rewards/repeat_penalty/std': 0.005356864538043737, 'rewards/near_duplicate_penalty/mean': 0.998106062412262, 'rewards/near_duplicate_penalty/std': 0.005356864538043737, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6377477645874023, 'rewards/total_composite/std': 0.3818265199661255, 'reward': 0.6377477645874023, 'reward_std': 0.3818265199661255, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1056709885597229, 'sampling/sampling_logp_difference/max': 2.4557881355285645, 'sampling/importance_sampling_ratio/min': 0.08579555153846741, 'sampling/importance_sampling_ratio/mean': 0.9991392493247986, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3120129220187664, 'clip_ratio/low_mean': 0.010416666977107525, 'clip_ratio/low_min': 0.010416666977107525, 'clip_ratio/high_mean': 0.06031998083926737, 'clip_ratio/high_max': 0.06031998083926737, 'clip_ratio/region_mean': 0.0707366478163749, 'reward_total_mean': 0.6377477645874023, 'reward_meter_mean': 0.6927865743637085, 'reward_meter_std': 0.3873361349105835, '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_lexical_plausibility_mean': 0.9700043797492981, 'reward_lexical_plausibility_std': 0.08484045416116714, 'reward_repeat_penalty_mean': 0.998106062412262, 'reward_repeat_penalty_std': 0.005356864538043737, 'reward_near_duplicate_penalty_mean': 0.998106062412262, 'reward_near_duplicate_penalty_std': 0.005356864538043737, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7637499570846558, 'reward_judge_quality_std': 0.3349173665046692, 'reward_total_composite_mean': 0.6377477645874023, 'reward_total_composite_std': 0.3818265199661255, 'epoch': 0.03}

 25%|██▌       | 830/3300 [3:18:31<8:00:21, 11.67s/it]INFO 04-12 14:04:44 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 831/3300 [3:18:45<8:23:38, 12.24s/it]2026-04-12 14:04:57,565 | INFO | train_grpo_train | metrics_logged mode=train step=831

                                                      
{'loss': -0.0925, 'grad_norm': 3.299501895904541, 'learning_rate': 7.484848484848486e-06, 'num_tokens': 1690215.0, 'completions/mean_length': 243.125, 'completions/min_length': 75.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 81.80000305175781, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.2392055094242096, 'rewards/meter/std': 0.28342151641845703, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.2519763112068176, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.9162173271179199, 'rewards/lexical_plausibility/std': 0.12281563133001328, 'rewards/judge_quality/mean': 0.23750001192092896, 'rewards/judge_quality/std': 0.15526476502418518, 'rewards/repeat_penalty/mean': 0.877632737159729, 'rewards/repeat_penalty/std': 0.0954660102725029, 'rewards/near_duplicate_penalty/mean': 0.9493685364723206, 'rewards/near_duplicate_penalty/std': 0.05328753963112831, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9554908275604248, 'rewards/distinct_2/std': 0.04688442870974541, 'rewards/total_composite/mean': 0.0472441241145134, 'rewards/total_composite/std': 0.0789516270160675, 'reward': 0.0472441241145134, 'reward_std': 0.0789516270160675, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16318044066429138, 'sampling/sampling_logp_difference/max': 2.1976687908172607, 'sampling/importance_sampling_ratio/min': 0.11106175929307938, 'sampling/importance_sampling_ratio/mean': 1.0221236944198608, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6983982473611832, 'clip_ratio/low_mean': 0.04283104743808508, 'clip_ratio/low_min': 0.04283104743808508, 'clip_ratio/high_mean': 0.056785713881254196, 'clip_ratio/high_max': 0.056785713881254196, 'clip_ratio/region_mean': 0.09961676131933928, 'reward_total_mean': 0.0472441241145134, 'reward_meter_mean': 0.2392055094242096, 'reward_meter_std': 0.28342151641845703, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.2519763112068176, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.9162173271179199, 'reward_lexical_plausibility_std': 0.12281563133001328, 'reward_repeat_penalty_mean': 0.877632737159729, 'reward_repeat_penalty_std': 0.0954660102725029, 'reward_near_duplicate_penalty_mean': 0.9493685364723206, 'reward_near_duplicate_penalty_std': 0.05328753963112831, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9554908275604248, 'reward_distinct_2_std': 0.04688442870974541, 'reward_judge_quality_mean': 0.23750001192092896, 'reward_judge_quality_std': 0.15526476502418518, 'reward_total_composite_mean': 0.0472441241145134, 'reward_total_composite_std': 0.0789516270160675, 'epoch': 0.03}

 25%|██▌       | 831/3300 [3:18:45<8:23:38, 12.24s/it]INFO 04-12 14:04:57 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 832/3300 [3:19:00<8:54:46, 13.00s/it]2026-04-12 14:05:12,340 | INFO | train_grpo_train | metrics_logged mode=train step=832

                                                      
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 7.481818181818182e-06, 'num_tokens': 1692071.0, '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': 0.6423313617706299, 'rewards/meter/std': 0.4211899936199188, 'rewards/count_adherence/mean': 0.9411764740943909, 'rewards/count_adherence/std': 0.06288499385118484, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 0.6244285106658936, 'rewards/near_duplicate_penalty/std': 0.13046982884407043, 'rewards/opening_diversity/mean': 0.7550863027572632, 'rewards/opening_diversity/std': 0.2533925771713257, 'rewards/distinct_2/mean': 0.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.08885331451892853, 'rewards/total_composite/std': 0.05813891440629959, 'reward': 0.08885331451892853, 'reward_std': 0.05813891440629959, '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': 0.08885331451892853, 'reward_meter_mean': 0.6423313617706299, 'reward_meter_std': 0.4211899936199188, 'reward_count_adherence_mean': 0.9411764740943909, 'reward_count_adherence_std': 0.06288499385118484, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.0, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 0.6244285106658936, 'reward_near_duplicate_penalty_std': 0.13046982884407043, 'reward_opening_diversity_mean': 0.7550863027572632, 'reward_opening_diversity_std': 0.2533925771713257, 'reward_distinct_2_mean': 0.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.08885331451892853, 'reward_total_composite_std': 0.05813891440629959, 'epoch': 0.03}

 25%|██▌       | 832/3300 [3:19:00<8:54:46, 13.00s/it]INFO 04-12 14:05:12 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 833/3300 [3:19:13<9:01:04, 13.16s/it]2026-04-12 14:05:25,872 | INFO | train_grpo_train | metrics_logged mode=train step=833

                                                      
{'loss': -0.0261, 'grad_norm': 3.6566264629364014, 'learning_rate': 7.47878787878788e-06, 'num_tokens': 1693447.0, 'completions/mean_length': 147.0, 'completions/min_length': 23.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 25.33333396911621, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.5361067056655884, 'rewards/meter/std': 0.39208167791366577, '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/lexical_plausibility/mean': 0.9393448233604431, 'rewards/lexical_plausibility/std': 0.11451318860054016, 'rewards/judge_quality/mean': 0.5137500166893005, 'rewards/judge_quality/std': 0.3703642189502716, 'rewards/repeat_penalty/mean': 0.7643716335296631, 'rewards/repeat_penalty/std': 0.10628793388605118, 'rewards/near_duplicate_penalty/mean': 0.977495551109314, 'rewards/near_duplicate_penalty/std': 0.06365221738815308, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2797941267490387, 'rewards/total_composite/std': 0.3050786256790161, 'reward': 0.2797941267490387, 'reward_std': 0.3050785958766937, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11691392213106155, 'sampling/sampling_logp_difference/max': 1.3519599437713623, 'sampling/importance_sampling_ratio/min': 0.2587326765060425, 'sampling/importance_sampling_ratio/mean': 1.0241702795028687, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.46255865320563316, 'clip_ratio/low_mean': 0.045495014637708664, 'clip_ratio/low_min': 0.045495014637708664, 'clip_ratio/high_mean': 0.03691123239696026, 'clip_ratio/high_max': 0.03691123239696026, 'clip_ratio/region_mean': 0.08240624703466892, 'reward_total_mean': 0.2797941267490387, 'reward_meter_mean': 0.5361067056655884, 'reward_meter_std': 0.39208167791366577, '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_lexical_plausibility_mean': 0.9393448233604431, 'reward_lexical_plausibility_std': 0.11451318860054016, 'reward_repeat_penalty_mean': 0.7643716335296631, 'reward_repeat_penalty_std': 0.10628793388605118, 'reward_near_duplicate_penalty_mean': 0.977495551109314, 'reward_near_duplicate_penalty_std': 0.06365221738815308, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5137500166893005, 'reward_judge_quality_std': 0.3703642189502716, 'reward_total_composite_mean': 0.2797941267490387, 'reward_total_composite_std': 0.3050786256790161, 'epoch': 0.03}

 25%|██▌       | 833/3300 [3:19:13<9:01:04, 13.16s/it]INFO 04-12 14:05:26 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 834/3300 [3:19:28<9:17:55, 13.57s/it]2026-04-12 14:05:40,427 | INFO | train_grpo_train | metrics_logged mode=train step=834

                                                      
{'loss': -0.0338, 'grad_norm': 1.1344773769378662, 'learning_rate': 7.4757575757575765e-06, 'num_tokens': 1694730.0, 'completions/mean_length': 454.375, 'completions/min_length': 51.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.875, 'completions/mean_terminated_length': 51.0, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.3278172016143799, 'rewards/meter/std': 0.35006678104400635, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 0.125, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.8194913268089294, 'rewards/lexical_plausibility/std': 0.11611448973417282, 'rewards/judge_quality/mean': 0.1587500125169754, 'rewards/judge_quality/std': 0.3075914680957794, 'rewards/repeat_penalty/mean': 0.9270184636116028, 'rewards/repeat_penalty/std': 0.11105029284954071, 'rewards/near_duplicate_penalty/mean': 0.9910468459129333, 'rewards/near_duplicate_penalty/std': 0.019919803366065025, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.998443067073822, 'rewards/distinct_2/std': 0.0031093021389096975, 'rewards/total_composite/mean': 0.028080236166715622, 'rewards/total_composite/std': 0.07942290604114532, 'reward': 0.028080236166715622, 'reward_std': 0.07942289859056473, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13641683757305145, 'sampling/sampling_logp_difference/max': 1.346531867980957, 'sampling/importance_sampling_ratio/min': 0.26014089584350586, 'sampling/importance_sampling_ratio/mean': 0.9992741942405701, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.0773949921131134, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.014705882407724857, 'clip_ratio/high_max': 0.014705882407724857, 'clip_ratio/region_mean': 0.014705882407724857, 'reward_total_mean': 0.028080236166715622, 'reward_meter_mean': 0.3278172016143799, 'reward_meter_std': 0.35006678104400635, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 0.125, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.8194913268089294, 'reward_lexical_plausibility_std': 0.11611448973417282, 'reward_repeat_penalty_mean': 0.9270184636116028, 'reward_repeat_penalty_std': 0.11105029284954071, 'reward_near_duplicate_penalty_mean': 0.9910468459129333, 'reward_near_duplicate_penalty_std': 0.019919803366065025, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.998443067073822, 'reward_distinct_2_std': 0.0031093021389096975, 'reward_judge_quality_mean': 0.1587500125169754, 'reward_judge_quality_std': 0.3075914680957794, 'reward_total_composite_mean': 0.028080236166715622, 'reward_total_composite_std': 0.07942290604114532, 'epoch': 0.03}

 25%|██▌       | 834/3300 [3:19:28<9:17:55, 13.57s/it]INFO 04-12 14:05:40 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 835/3300 [3:19:42<9:21:05, 13.66s/it]2026-04-12 14:05:54,266 | INFO | train_grpo_train | metrics_logged mode=train step=835

                                                      
{'loss': -0.029, 'grad_norm': 5.611277103424072, 'learning_rate': 7.472727272727274e-06, 'num_tokens': 1697034.0, 'completions/mean_length': 184.0, 'completions/min_length': 112.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 137.1428680419922, 'completions/min_terminated_length': 112.0, 'completions/max_terminated_length': 195.0, 'rewards/meter/mean': 0.7396829128265381, 'rewards/meter/std': 0.3543570339679718, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.1414213478565216, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9472166299819946, 'rewards/lexical_plausibility/std': 0.10670626163482666, 'rewards/judge_quality/mean': 0.1875, 'rewards/judge_quality/std': 0.1060660183429718, 'rewards/repeat_penalty/mean': 0.41893428564071655, 'rewards/repeat_penalty/std': 0.36091017723083496, 'rewards/near_duplicate_penalty/mean': 0.8077979683876038, 'rewards/near_duplicate_penalty/std': 0.12983903288841248, 'rewards/opening_diversity/mean': 0.9124999642372131, 'rewards/opening_diversity/std': 0.13562026619911194, 'rewards/distinct_2/mean': 0.5231683850288391, 'rewards/distinct_2/std': 0.3322204351425171, 'rewards/total_composite/mean': 0.12948951125144958, 'rewards/total_composite/std': 0.1072273850440979, 'reward': 0.12948951125144958, 'reward_std': 0.1072273850440979, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09975112974643707, 'sampling/sampling_logp_difference/max': 2.4648022651672363, 'sampling/importance_sampling_ratio/min': 0.08502565324306488, 'sampling/importance_sampling_ratio/mean': 1.0083603858947754, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6093812212347984, 'clip_ratio/low_mean': 0.02280219830572605, 'clip_ratio/low_min': 0.02280219830572605, 'clip_ratio/high_mean': 0.055024812929332256, 'clip_ratio/high_max': 0.055024812929332256, 'clip_ratio/region_mean': 0.07782701123505831, 'reward_total_mean': 0.12948951125144958, 'reward_meter_mean': 0.7396829128265381, 'reward_meter_std': 0.3543570339679718, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.1414213478565216, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9472166299819946, 'reward_lexical_plausibility_std': 0.10670626163482666, 'reward_repeat_penalty_mean': 0.41893428564071655, 'reward_repeat_penalty_std': 0.36091017723083496, 'reward_near_duplicate_penalty_mean': 0.8077979683876038, 'reward_near_duplicate_penalty_std': 0.12983903288841248, 'reward_opening_diversity_mean': 0.9124999642372131, 'reward_opening_diversity_std': 0.13562026619911194, 'reward_distinct_2_mean': 0.5231683850288391, 'reward_distinct_2_std': 0.3322204351425171, 'reward_judge_quality_mean': 0.1875, 'reward_judge_quality_std': 0.1060660183429718, 'reward_total_composite_mean': 0.12948951125144958, 'reward_total_composite_std': 0.1072273850440979, 'epoch': 0.03}

 25%|██▌       | 835/3300 [3:19:42<9:21:05, 13.66s/it]INFO 04-12 14:05:54 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 836/3300 [3:19:55<9:21:07, 13.66s/it]2026-04-12 14:06:07,945 | INFO | train_grpo_train | metrics_logged mode=train step=836

                                                      
{'loss': -0.1061, 'grad_norm': 5.153656959533691, 'learning_rate': 7.46969696969697e-06, 'num_tokens': 1699147.0, 'completions/mean_length': 153.125, 'completions/min_length': 87.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 101.85714721679688, 'completions/min_terminated_length': 87.0, 'completions/max_terminated_length': 114.0, 'rewards/meter/mean': 0.7257121801376343, 'rewards/meter/std': 0.28514406085014343, '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/lexical_plausibility/mean': 0.9867037534713745, 'rewards/lexical_plausibility/std': 0.03760756924748421, 'rewards/judge_quality/mean': 0.39625000953674316, 'rewards/judge_quality/std': 0.222898930311203, 'rewards/repeat_penalty/mean': 0.6885099411010742, 'rewards/repeat_penalty/std': 0.2579411566257477, 'rewards/near_duplicate_penalty/mean': 0.9214573502540588, 'rewards/near_duplicate_penalty/std': 0.08484853059053421, 'rewards/opening_diversity/mean': 0.828125, 'rewards/opening_diversity/std': 0.17912860214710236, 'rewards/distinct_2/mean': 0.86307692527771, 'rewards/distinct_2/std': 0.10983160138130188, 'rewards/total_composite/mean': 0.25582242012023926, 'rewards/total_composite/std': 0.20269383490085602, 'reward': 0.25582242012023926, 'reward_std': 0.20269383490085602, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12152904272079468, 'sampling/sampling_logp_difference/max': 1.9659693241119385, 'sampling/importance_sampling_ratio/min': 0.14002008736133575, 'sampling/importance_sampling_ratio/mean': 1.0027951002120972, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.46619174629449844, 'clip_ratio/low_mean': 0.0416061719879508, 'clip_ratio/low_min': 0.0416061719879508, 'clip_ratio/high_mean': 0.0638554748147726, 'clip_ratio/high_max': 0.0638554748147726, 'clip_ratio/region_mean': 0.10546164680272341, 'reward_total_mean': 0.25582242012023926, 'reward_meter_mean': 0.7257121801376343, 'reward_meter_std': 0.28514406085014343, '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_lexical_plausibility_mean': 0.9867037534713745, 'reward_lexical_plausibility_std': 0.03760756924748421, 'reward_repeat_penalty_mean': 0.6885099411010742, 'reward_repeat_penalty_std': 0.2579411566257477, 'reward_near_duplicate_penalty_mean': 0.9214573502540588, 'reward_near_duplicate_penalty_std': 0.08484853059053421, 'reward_opening_diversity_mean': 0.828125, 'reward_opening_diversity_std': 0.17912860214710236, 'reward_distinct_2_mean': 0.86307692527771, 'reward_distinct_2_std': 0.10983160138130188, 'reward_judge_quality_mean': 0.39625000953674316, 'reward_judge_quality_std': 0.222898930311203, 'reward_total_composite_mean': 0.25582242012023926, 'reward_total_composite_std': 0.20269383490085602, 'epoch': 0.03}

 25%|██▌       | 836/3300 [3:19:55<9:21:07, 13.66s/it]INFO 04-12 14:06:08 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 837/3300 [3:20:08<9:13:53, 13.49s/it]2026-04-12 14:06:21,035 | INFO | train_grpo_train | metrics_logged mode=train step=837

                                                      
{'loss': -0.0394, 'grad_norm': 1.0968133211135864, 'learning_rate': 7.4666666666666675e-06, 'num_tokens': 1700414.0, 'completions/mean_length': 205.375, 'completions/min_length': 19.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 21.399999618530273, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.6317184567451477, 'rewards/meter/std': 0.48563364148139954, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.927947461605072, 'rewards/lexical_plausibility/std': 0.10021715611219406, 'rewards/judge_quality/mean': 0.5012500286102295, 'rewards/judge_quality/std': 0.4569600820541382, 'rewards/repeat_penalty/mean': 0.693365752696991, 'rewards/repeat_penalty/std': 0.1168353334069252, 'rewards/near_duplicate_penalty/mean': 0.939164400100708, 'rewards/near_duplicate_penalty/std': 0.1186617836356163, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9786324501037598, 'rewards/distinct_2/std': 0.06043647602200508, 'rewards/total_composite/mean': 0.4740714728832245, 'rewards/total_composite/std': 0.4698944389820099, 'reward': 0.4740714728832245, 'reward_std': 0.4698944091796875, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08677541464567184, 'sampling/sampling_logp_difference/max': 1.6270151138305664, 'sampling/importance_sampling_ratio/min': 0.1965152770280838, 'sampling/importance_sampling_ratio/mean': 1.0064005851745605, 'sampling/importance_sampling_ratio/max': 1.5441728830337524, 'entropy': 0.3056196980178356, 'clip_ratio/low_mean': 0.01785714365541935, 'clip_ratio/low_min': 0.01785714365541935, 'clip_ratio/high_mean': 0.044728299137204885, 'clip_ratio/high_max': 0.044728299137204885, 'clip_ratio/region_mean': 0.06258544279262424, 'reward_total_mean': 0.4740714728832245, 'reward_meter_mean': 0.6317184567451477, 'reward_meter_std': 0.48563364148139954, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.927947461605072, 'reward_lexical_plausibility_std': 0.10021715611219406, 'reward_repeat_penalty_mean': 0.693365752696991, 'reward_repeat_penalty_std': 0.1168353334069252, 'reward_near_duplicate_penalty_mean': 0.939164400100708, 'reward_near_duplicate_penalty_std': 0.1186617836356163, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9786324501037598, 'reward_distinct_2_std': 0.06043647602200508, 'reward_judge_quality_mean': 0.5012500286102295, 'reward_judge_quality_std': 0.4569600820541382, 'reward_total_composite_mean': 0.4740714728832245, 'reward_total_composite_std': 0.4698944389820099, 'epoch': 0.03}

 25%|██▌       | 837/3300 [3:20:09<9:13:53, 13.49s/it]INFO 04-12 14:06:21 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 838/3300 [3:20:23<9:29:09, 13.87s/it]2026-04-12 14:06:35,791 | INFO | train_grpo_train | metrics_logged mode=train step=838

                                                      
{'loss': -0.1768, 'grad_norm': 3.3862082958221436, 'learning_rate': 7.463636363636364e-06, 'num_tokens': 1703069.0, 'completions/mean_length': 306.875, 'completions/min_length': 168.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 183.8000030517578, 'completions/min_terminated_length': 168.0, 'completions/max_terminated_length': 208.0, 'rewards/meter/mean': 0.39931076765060425, 'rewards/meter/std': 0.27909573912620544, 'rewards/count_adherence/mean': 0.703125, 'rewards/count_adherence/std': 0.3269003629684448, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9378780126571655, 'rewards/lexical_plausibility/std': 0.10751595348119736, 'rewards/judge_quality/mean': 0.20000000298023224, 'rewards/judge_quality/std': 0.1414213478565216, 'rewards/repeat_penalty/mean': 0.6501808166503906, 'rewards/repeat_penalty/std': 0.3393625020980835, 'rewards/near_duplicate_penalty/mean': 0.8912743926048279, 'rewards/near_duplicate_penalty/std': 0.13243815302848816, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.7067744135856628, 'rewards/distinct_2/std': 0.3037939667701721, 'rewards/total_composite/mean': 0.091504767537117, 'rewards/total_composite/std': 0.09957390278577805, 'reward': 0.091504767537117, 'reward_std': 0.09957390278577805, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1197218969464302, 'sampling/sampling_logp_difference/max': 2.356180191040039, 'sampling/importance_sampling_ratio/min': 0.09478157758712769, 'sampling/importance_sampling_ratio/mean': 1.0022670030593872, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3877883516252041, 'clip_ratio/low_mean': 0.015739469090476632, 'clip_ratio/low_min': 0.015739469090476632, 'clip_ratio/high_mean': 0.04271893482655287, 'clip_ratio/high_max': 0.04271893482655287, 'clip_ratio/region_mean': 0.0584584039170295, 'reward_total_mean': 0.091504767537117, 'reward_meter_mean': 0.39931076765060425, 'reward_meter_std': 0.27909573912620544, 'reward_count_adherence_mean': 0.703125, 'reward_count_adherence_std': 0.3269003629684448, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9378780126571655, 'reward_lexical_plausibility_std': 0.10751595348119736, 'reward_repeat_penalty_mean': 0.6501808166503906, 'reward_repeat_penalty_std': 0.3393625020980835, 'reward_near_duplicate_penalty_mean': 0.8912743926048279, 'reward_near_duplicate_penalty_std': 0.13243815302848816, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.7067744135856628, 'reward_distinct_2_std': 0.3037939667701721, 'reward_judge_quality_mean': 0.20000000298023224, 'reward_judge_quality_std': 0.1414213478565216, 'reward_total_composite_mean': 0.091504767537117, 'reward_total_composite_std': 0.09957390278577805, 'epoch': 0.03}

 25%|██▌       | 838/3300 [3:20:23<9:29:09, 13.87s/it]INFO 04-12 14:06:36 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 839/3300 [3:20:37<9:27:38, 13.84s/it]2026-04-12 14:06:49,554 | INFO | train_grpo_train | metrics_logged mode=train step=839

                                                      
{'loss': -0.0879, 'grad_norm': 5.077421188354492, 'learning_rate': 7.460606060606061e-06, 'num_tokens': 1704950.0, 'completions/mean_length': 134.125, 'completions/min_length': 71.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 80.14286041259766, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 92.0, 'rewards/meter/mean': 0.808836817741394, 'rewards/meter/std': 0.18064318597316742, '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/lexical_plausibility/mean': 0.9653936624526978, 'rewards/lexical_plausibility/std': 0.09788139909505844, 'rewards/judge_quality/mean': 0.5287500023841858, 'rewards/judge_quality/std': 0.3231734335422516, 'rewards/repeat_penalty/mean': 0.8903459310531616, 'rewards/repeat_penalty/std': 0.15989521145820618, 'rewards/near_duplicate_penalty/mean': 0.957159161567688, 'rewards/near_duplicate_penalty/std': 0.0361555740237236, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.945880115032196, 'rewards/distinct_2/std': 0.0991036668419838, 'rewards/total_composite/mean': 0.41259628534317017, 'rewards/total_composite/std': 0.3446335196495056, 'reward': 0.41259628534317017, 'reward_std': 0.3446335196495056, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1046198308467865, 'sampling/sampling_logp_difference/max': 2.309542655944824, 'sampling/importance_sampling_ratio/min': 0.09930665791034698, 'sampling/importance_sampling_ratio/mean': 1.0194816589355469, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5719370767474174, 'clip_ratio/low_mean': 0.04114250931888819, 'clip_ratio/low_min': 0.04114250931888819, 'clip_ratio/high_mean': 0.052555736154317856, 'clip_ratio/high_max': 0.052555736154317856, 'clip_ratio/region_mean': 0.09369824547320604, 'reward_total_mean': 0.41259628534317017, 'reward_meter_mean': 0.808836817741394, 'reward_meter_std': 0.18064318597316742, '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_lexical_plausibility_mean': 0.9653936624526978, 'reward_lexical_plausibility_std': 0.09788139909505844, 'reward_repeat_penalty_mean': 0.8903459310531616, 'reward_repeat_penalty_std': 0.15989521145820618, 'reward_near_duplicate_penalty_mean': 0.957159161567688, 'reward_near_duplicate_penalty_std': 0.0361555740237236, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.945880115032196, 'reward_distinct_2_std': 0.0991036668419838, 'reward_judge_quality_mean': 0.5287500023841858, 'reward_judge_quality_std': 0.3231734335422516, 'reward_total_composite_mean': 0.41259628534317017, 'reward_total_composite_std': 0.3446335196495056, 'epoch': 0.03}

 25%|██▌       | 839/3300 [3:20:37<9:27:38, 13.84s/it]INFO 04-12 14:06:49 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 840/3300 [3:20:50<9:22:46, 13.73s/it]2026-04-12 14:07:03,023 | INFO | train_grpo_train | metrics_logged mode=train step=840

                                                      
{'loss': -0.0255, 'grad_norm': 0.8306662440299988, 'learning_rate': 7.4575757575757575e-06, 'num_tokens': 1706145.0, 'completions/mean_length': 390.375, 'completions/min_length': 24.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.75, 'completions/mean_terminated_length': 25.5, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.2271895855665207, 'rewards/meter/std': 0.408864289522171, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.25, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.7962422370910645, 'rewards/lexical_plausibility/std': 0.14310558140277863, 'rewards/judge_quality/mean': 0.26875001192092896, 'rewards/judge_quality/std': 0.4208303987979889, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.21006235480308533, 'rewards/total_composite/std': 0.3916552662849426, 'reward': 0.21006235480308533, 'reward_std': 0.39165523648262024, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10430178046226501, 'sampling/sampling_logp_difference/max': 0.7973198890686035, 'sampling/importance_sampling_ratio/min': 0.450534850358963, 'sampling/importance_sampling_ratio/mean': 1.0402921438217163, 'sampling/importance_sampling_ratio/max': 1.7328897714614868, 'entropy': 0.14956891536712646, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.014467592816799879, 'clip_ratio/high_max': 0.014467592816799879, 'clip_ratio/region_mean': 0.014467592816799879, 'reward_total_mean': 0.21006235480308533, 'reward_meter_mean': 0.2271895855665207, 'reward_meter_std': 0.408864289522171, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.25, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.7962422370910645, 'reward_lexical_plausibility_std': 0.14310558140277863, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.26875001192092896, 'reward_judge_quality_std': 0.4208303987979889, 'reward_total_composite_mean': 0.21006235480308533, 'reward_total_composite_std': 0.3916552662849426, 'epoch': 0.03}

 25%|██▌       | 840/3300 [3:20:50<9:22:46, 13.73s/it]INFO 04-12 14:07:03 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 841/3300 [3:21:04<9:25:13, 13.79s/it]2026-04-12 14:07:16,976 | INFO | train_grpo_train | metrics_logged mode=train step=841

                                                      
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 7.454545454545456e-06, 'num_tokens': 1707497.0, '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': 0.015986643731594086, 'rewards/meter/std': 0.022730566561222076, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.3720119297504425, 'rewards/arabic_clean/mean': 0.25, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.7878352403640747, 'rewards/lexical_plausibility/std': 0.04710675776004791, 'rewards/judge_quality/mean': 0.05000000074505806, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.9990274906158447, 'rewards/repeat_penalty/std': 0.0018006979953497648, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9990274906158447, 'rewards/distinct_2/std': 0.0018006979953497648, 'rewards/total_composite/mean': 2.4907198167056777e-05, 'rewards/total_composite/std': 7.044819358270615e-05, 'reward': 2.4907198167056777e-05, 'reward_std': 7.044819358270615e-05, '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': 2.4907198167056777e-05, 'reward_meter_mean': 0.015986643731594086, 'reward_meter_std': 0.022730566561222076, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.3720119297504425, 'reward_arabic_clean_mean': 0.25, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.7878352403640747, 'reward_lexical_plausibility_std': 0.04710675776004791, 'reward_repeat_penalty_mean': 0.9990274906158447, 'reward_repeat_penalty_std': 0.0018006979953497648, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9990274906158447, 'reward_distinct_2_std': 0.0018006979953497648, 'reward_judge_quality_mean': 0.05000000074505806, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 2.4907198167056777e-05, 'reward_total_composite_std': 7.044819358270615e-05, 'epoch': 0.03}

 25%|██▌       | 841/3300 [3:21:04<9:25:13, 13.79s/it]INFO 04-12 14:07:17 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 842/3300 [3:21:19<9:31:11, 13.94s/it]2026-04-12 14:07:31,265 | INFO | train_grpo_train | metrics_logged mode=train step=842

                                                      
{'loss': -0.0362, 'grad_norm': 1.4071052074432373, 'learning_rate': 7.451515151515152e-06, 'num_tokens': 1708809.0, 'completions/mean_length': 269.0, 'completions/min_length': 26.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 26.0, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.5345239043235779, 'rewards/meter/std': 0.4528908133506775, '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/lexical_plausibility/mean': 0.8868546485900879, 'rewards/lexical_plausibility/std': 0.12396898120641708, 'rewards/judge_quality/mean': 0.38875001668930054, 'rewards/judge_quality/std': 0.44121384620666504, 'rewards/repeat_penalty/mean': 0.7730461359024048, 'rewards/repeat_penalty/std': 0.09359873831272125, 'rewards/near_duplicate_penalty/mean': 0.9890615344047546, 'rewards/near_duplicate_penalty/std': 0.025376567617058754, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3438541889190674, 'rewards/total_composite/std': 0.44879963994026184, 'reward': 0.3438541889190674, 'reward_std': 0.44879963994026184, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12024195492267609, 'sampling/sampling_logp_difference/max': 0.9006625413894653, 'sampling/importance_sampling_ratio/min': 0.4856184720993042, 'sampling/importance_sampling_ratio/mean': 1.0317338705062866, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5495957285165787, 'clip_ratio/low_mean': 0.01923076994717121, 'clip_ratio/low_min': 0.01923076994717121, 'clip_ratio/high_mean': 0.024038461968302727, 'clip_ratio/high_max': 0.024038461968302727, 'clip_ratio/region_mean': 0.04326923191547394, 'reward_total_mean': 0.3438541889190674, 'reward_meter_mean': 0.5345239043235779, 'reward_meter_std': 0.4528908133506775, '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_lexical_plausibility_mean': 0.8868546485900879, 'reward_lexical_plausibility_std': 0.12396898120641708, 'reward_repeat_penalty_mean': 0.7730461359024048, 'reward_repeat_penalty_std': 0.09359873831272125, 'reward_near_duplicate_penalty_mean': 0.9890615344047546, 'reward_near_duplicate_penalty_std': 0.025376567617058754, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.38875001668930054, 'reward_judge_quality_std': 0.44121384620666504, 'reward_total_composite_mean': 0.3438541889190674, 'reward_total_composite_std': 0.44879963994026184, 'epoch': 0.03}

 26%|██▌       | 842/3300 [3:21:19<9:31:11, 13.94s/it]INFO 04-12 14:07:31 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 843/3300 [3:21:34<9:53:28, 14.49s/it]2026-04-12 14:07:47,040 | INFO | train_grpo_train | metrics_logged mode=train step=843

                                                      
{'loss': -0.3255, 'grad_norm': 2.028681755065918, 'learning_rate': 7.448484848484849e-06, 'num_tokens': 1713231.0, 'completions/mean_length': 406.75, 'completions/min_length': 314.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 371.66668701171875, 'completions/min_terminated_length': 314.0, 'completions/max_terminated_length': 442.0, 'rewards/meter/mean': 0.5781040787696838, 'rewards/meter/std': 0.16782401502132416, 'rewards/count_adherence/mean': 0.7291666865348816, 'rewards/count_adherence/std': 0.2878147065639496, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9710139632225037, 'rewards/lexical_plausibility/std': 0.08198493719100952, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.11649647355079651, 'rewards/repeat_penalty/mean': 0.5036746859550476, 'rewards/repeat_penalty/std': 0.32200655341148376, 'rewards/near_duplicate_penalty/mean': 0.9133760929107666, 'rewards/near_duplicate_penalty/std': 0.08105232566595078, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.5551795363426208, 'rewards/distinct_2/std': 0.3133257031440735, 'rewards/total_composite/mean': 0.13722610473632812, 'rewards/total_composite/std': 0.07641561329364777, 'reward': 0.13722610473632812, 'reward_std': 0.07641561329364777, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0807938203215599, 'sampling/sampling_logp_difference/max': 2.5774078369140625, 'sampling/importance_sampling_ratio/min': 0.07597067207098007, 'sampling/importance_sampling_ratio/mean': 1.0048794746398926, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3612729646265507, 'clip_ratio/low_mean': 0.006787330377846956, 'clip_ratio/low_min': 0.006787330377846956, 'clip_ratio/high_mean': 0.051659686490893364, 'clip_ratio/high_max': 0.051659686490893364, 'clip_ratio/region_mean': 0.05844701686874032, 'reward_total_mean': 0.13722610473632812, 'reward_meter_mean': 0.5781040787696838, 'reward_meter_std': 0.16782401502132416, 'reward_count_adherence_mean': 0.7291666865348816, 'reward_count_adherence_std': 0.2878147065639496, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9710139632225037, 'reward_lexical_plausibility_std': 0.08198493719100952, 'reward_repeat_penalty_mean': 0.5036746859550476, 'reward_repeat_penalty_std': 0.32200655341148376, 'reward_near_duplicate_penalty_mean': 0.9133760929107666, 'reward_near_duplicate_penalty_std': 0.08105232566595078, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.5551795363426208, 'reward_distinct_2_std': 0.3133257031440735, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.11649647355079651, 'reward_total_composite_mean': 0.13722610473632812, 'reward_total_composite_std': 0.07641561329364777, 'epoch': 0.03}

 26%|██▌       | 843/3300 [3:21:35<9:53:28, 14.49s/it]INFO 04-12 14:07:47 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 844/3300 [3:21:48<9:46:17, 14.32s/it]2026-04-12 14:08:00,971 | INFO | train_grpo_train | metrics_logged mode=train step=844

                                                      
{'loss': -0.0878, 'grad_norm': 4.211852550506592, 'learning_rate': 7.445454545454546e-06, 'num_tokens': 1715558.0, 'completions/mean_length': 152.875, 'completions/min_length': 91.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 101.5714340209961, 'completions/min_terminated_length': 91.0, 'completions/max_terminated_length': 108.0, 'rewards/meter/mean': 0.474073588848114, 'rewards/meter/std': 0.35706835985183716, '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/lexical_plausibility/mean': 0.9775367975234985, 'rewards/lexical_plausibility/std': 0.06353551149368286, 'rewards/judge_quality/mean': 0.20000000298023224, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.9689923524856567, 'rewards/repeat_penalty/std': 0.05518259108066559, 'rewards/near_duplicate_penalty/mean': 0.9857252836227417, 'rewards/near_duplicate_penalty/std': 0.009722024202346802, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.9898785352706909, 'rewards/distinct_2/std': 0.028627805411815643, 'rewards/total_composite/mean': 0.08947252482175827, 'rewards/total_composite/std': 0.06534622609615326, 'reward': 0.08947252482175827, 'reward_std': 0.06534622609615326, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10866747796535492, 'sampling/sampling_logp_difference/max': 1.9588477611541748, 'sampling/importance_sampling_ratio/min': 0.14102081954479218, 'sampling/importance_sampling_ratio/mean': 1.0037509202957153, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.45201993733644485, 'clip_ratio/low_mean': 0.040869816206395626, 'clip_ratio/low_min': 0.040869816206395626, 'clip_ratio/high_mean': 0.04959984961897135, 'clip_ratio/high_max': 0.04959984961897135, 'clip_ratio/region_mean': 0.09046966582536697, 'reward_total_mean': 0.08947252482175827, 'reward_meter_mean': 0.474073588848114, 'reward_meter_std': 0.35706835985183716, '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_lexical_plausibility_mean': 0.9775367975234985, 'reward_lexical_plausibility_std': 0.06353551149368286, 'reward_repeat_penalty_mean': 0.9689923524856567, 'reward_repeat_penalty_std': 0.05518259108066559, 'reward_near_duplicate_penalty_mean': 0.9857252836227417, 'reward_near_duplicate_penalty_std': 0.009722024202346802, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.9898785352706909, 'reward_distinct_2_std': 0.028627805411815643, 'reward_judge_quality_mean': 0.20000000298023224, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.08947252482175827, 'reward_total_composite_std': 0.06534622609615326, 'epoch': 0.03}

 26%|██▌       | 844/3300 [3:21:48<9:46:17, 14.32s/it]INFO 04-12 14:08:01 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 845/3300 [3:22:03<9:43:55, 14.27s/it]2026-04-12 14:08:15,118 | INFO | train_grpo_train | metrics_logged mode=train step=845

                                                      
{'loss': -0.047, 'grad_norm': 1.157021164894104, 'learning_rate': 7.442424242424243e-06, 'num_tokens': 1716911.0, 'completions/mean_length': 396.125, 'completions/min_length': 48.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.75, 'completions/mean_terminated_length': 48.5, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.6507343649864197, 'rewards/meter/std': 0.395896315574646, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.25, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.8463503122329712, 'rewards/lexical_plausibility/std': 0.11834544688463211, 'rewards/judge_quality/mean': 0.16249999403953552, 'rewards/judge_quality/std': 0.22320714592933655, 'rewards/repeat_penalty/mean': 0.954783022403717, 'rewards/repeat_penalty/std': 0.08504591137170792, 'rewards/near_duplicate_penalty/mean': 0.9879560470581055, 'rewards/near_duplicate_penalty/std': 0.018811138346791267, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9980278015136719, 'rewards/distinct_2/std': 0.003018119838088751, 'rewards/total_composite/mean': 0.0071954140439629555, 'rewards/total_composite/std': 0.01335498970001936, 'reward': 0.0071954140439629555, 'reward_std': 0.01335498783737421, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14272001385688782, 'sampling/sampling_logp_difference/max': 1.3716845512390137, 'sampling/importance_sampling_ratio/min': 0.2536792457103729, 'sampling/importance_sampling_ratio/mean': 0.9959360957145691, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.19408629834651947, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.04373937100172043, 'clip_ratio/high_max': 0.04373937100172043, 'clip_ratio/region_mean': 0.04373937100172043, 'reward_total_mean': 0.0071954140439629555, 'reward_meter_mean': 0.6507343649864197, 'reward_meter_std': 0.395896315574646, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.25, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.8463503122329712, 'reward_lexical_plausibility_std': 0.11834544688463211, 'reward_repeat_penalty_mean': 0.954783022403717, 'reward_repeat_penalty_std': 0.08504591137170792, 'reward_near_duplicate_penalty_mean': 0.9879560470581055, 'reward_near_duplicate_penalty_std': 0.018811138346791267, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9980278015136719, 'reward_distinct_2_std': 0.003018119838088751, 'reward_judge_quality_mean': 0.16249999403953552, 'reward_judge_quality_std': 0.22320714592933655, 'reward_total_composite_mean': 0.0071954140439629555, 'reward_total_composite_std': 0.01335498970001936, 'epoch': 0.03}

 26%|██▌       | 845/3300 [3:22:03<9:43:55, 14.27s/it]INFO 04-12 14:08:15 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 846/3300 [3:22:15<9:19:07, 13.67s/it]2026-04-12 14:08:27,379 | INFO | train_grpo_train | metrics_logged mode=train step=846

                                                      
{'loss': 0.0032, 'grad_norm': 4.8742594718933105, 'learning_rate': 7.439393939393939e-06, 'num_tokens': 1720998.0, 'completions/mean_length': 266.875, 'completions/min_length': 229.0, 'completions/max_length': 314.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 266.875, 'completions/min_terminated_length': 229.0, 'completions/max_terminated_length': 314.0, 'rewards/meter/mean': 0.5946990847587585, 'rewards/meter/std': 0.26137739419937134, 'rewards/count_adherence/mean': 0.9305555820465088, 'rewards/count_adherence/std': 0.10179170221090317, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.07440237700939178, 'rewards/repeat_penalty/mean': 0.3091259002685547, 'rewards/repeat_penalty/std': 0.2901867628097534, 'rewards/near_duplicate_penalty/mean': 0.8128618001937866, 'rewards/near_duplicate_penalty/std': 0.13602638244628906, 'rewards/opening_diversity/mean': 0.9382575750350952, 'rewards/opening_diversity/std': 0.07783199101686478, 'rewards/distinct_2/mean': 0.37824079394340515, 'rewards/distinct_2/std': 0.32145142555236816, 'rewards/total_composite/mean': 0.1485380232334137, 'rewards/total_composite/std': 0.10988743603229523, 'reward': 0.1485380232334137, 'reward_std': 0.10988743603229523, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07831436395645142, 'sampling/sampling_logp_difference/max': 3.982466697692871, 'sampling/importance_sampling_ratio/min': 0.01863960362970829, 'sampling/importance_sampling_ratio/mean': 1.0055218935012817, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3786616772413254, 'clip_ratio/low_mean': 0.029865495627745986, 'clip_ratio/low_min': 0.029865495627745986, 'clip_ratio/high_mean': 0.0433481028303504, 'clip_ratio/high_max': 0.0433481028303504, 'clip_ratio/region_mean': 0.07321359845809639, 'reward_total_mean': 0.1485380232334137, 'reward_meter_mean': 0.5946990847587585, 'reward_meter_std': 0.26137739419937134, 'reward_count_adherence_mean': 0.9305555820465088, 'reward_count_adherence_std': 0.10179170221090317, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.3091259002685547, 'reward_repeat_penalty_std': 0.2901867628097534, 'reward_near_duplicate_penalty_mean': 0.8128618001937866, 'reward_near_duplicate_penalty_std': 0.13602638244628906, 'reward_opening_diversity_mean': 0.9382575750350952, 'reward_opening_diversity_std': 0.07783199101686478, 'reward_distinct_2_mean': 0.37824079394340515, 'reward_distinct_2_std': 0.32145142555236816, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.07440237700939178, 'reward_total_composite_mean': 0.1485380232334137, 'reward_total_composite_std': 0.10988743603229523, 'epoch': 0.03}

 26%|██▌       | 846/3300 [3:22:15<9:19:07, 13.67s/it]INFO 04-12 14:08:27 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 847/3300 [3:22:29<9:30:43, 13.96s/it]2026-04-12 14:08:42,033 | INFO | train_grpo_train | metrics_logged mode=train step=847

                                                      
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 7.4363636363636375e-06, 'num_tokens': 1722398.0, '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': 0.07988806068897247, 'rewards/meter/std': 0.18823161721229553, 'rewards/count_adherence/mean': 0.6875, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 0.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.7785017490386963, 'rewards/lexical_plausibility/std': 0.030447019264101982, 'rewards/judge_quality/mean': 0.05000000074505806, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.90625, 'rewards/repeat_penalty/std': 0.12938730418682098, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.0, 'rewards/total_composite/std': 0.0, 'reward': 0.0, 'reward_std': 0.0, 'frac_reward_zero_std': 1.0, 'sampling/sampling_logp_difference/mean': 0.0, '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': 0.0, 'reward_meter_mean': 0.07988806068897247, 'reward_meter_std': 0.18823161721229553, 'reward_count_adherence_mean': 0.6875, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 0.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.7785017490386963, 'reward_lexical_plausibility_std': 0.030447019264101982, 'reward_repeat_penalty_mean': 0.90625, 'reward_repeat_penalty_std': 0.12938730418682098, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.05000000074505806, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.0, 'reward_total_composite_std': 0.0, 'epoch': 0.03}

 26%|██▌       | 847/3300 [3:22:30<9:30:43, 13.96s/it]INFO 04-12 14:08:42 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 848/3300 [3:22:40<8:43:13, 12.80s/it]2026-04-12 14:08:52,124 | INFO | train_grpo_train | metrics_logged mode=train step=848

                                                      
{'loss': 0.0161, 'grad_norm': 10.770121574401855, 'learning_rate': 7.433333333333334e-06, 'num_tokens': 1724479.0, 'completions/mean_length': 79.125, 'completions/min_length': 67.0, 'completions/max_length': 86.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.125, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.8448991775512695, 'rewards/meter/std': 0.21849018335342407, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.47874999046325684, 'rewards/judge_quality/std': 0.19208908081054688, 'rewards/repeat_penalty/mean': 0.8118500709533691, 'rewards/repeat_penalty/std': 0.19468837976455688, 'rewards/near_duplicate_penalty/mean': 0.9453835487365723, 'rewards/near_duplicate_penalty/std': 0.05366896465420723, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8872851133346558, 'rewards/distinct_2/std': 0.10600007325410843, 'rewards/total_composite/mean': 0.4133495092391968, 'rewards/total_composite/std': 0.22245998680591583, 'reward': 0.4133495092391968, 'reward_std': 0.22245998680591583, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11687285453081131, 'sampling/sampling_logp_difference/max': 1.5692310333251953, 'sampling/importance_sampling_ratio/min': 0.2082052230834961, 'sampling/importance_sampling_ratio/mean': 1.0286939144134521, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7251513004302979, 'clip_ratio/low_mean': 0.07596874516457319, 'clip_ratio/low_min': 0.07596874516457319, 'clip_ratio/high_mean': 0.04243226069957018, 'clip_ratio/high_max': 0.04243226069957018, 'clip_ratio/region_mean': 0.11840100586414337, 'reward_total_mean': 0.4133495092391968, 'reward_meter_mean': 0.8448991775512695, 'reward_meter_std': 0.21849018335342407, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8118500709533691, 'reward_repeat_penalty_std': 0.19468837976455688, 'reward_near_duplicate_penalty_mean': 0.9453835487365723, 'reward_near_duplicate_penalty_std': 0.05366896465420723, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8872851133346558, 'reward_distinct_2_std': 0.10600007325410843, 'reward_judge_quality_mean': 0.47874999046325684, 'reward_judge_quality_std': 0.19208908081054688, 'reward_total_composite_mean': 0.4133495092391968, 'reward_total_composite_std': 0.22245998680591583, 'epoch': 0.03}

 26%|██▌       | 848/3300 [3:22:40<8:43:13, 12.80s/it]INFO 04-12 14:08:52 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 849/3300 [3:22:51<8:27:15, 12.42s/it]2026-04-12 14:09:03,642 | INFO | train_grpo_train | metrics_logged mode=train step=849

                                                      
{'loss': -0.0103, 'grad_norm': 7.369930744171143, 'learning_rate': 7.430303030303031e-06, 'num_tokens': 1727847.0, 'completions/mean_length': 203.0, 'completions/min_length': 185.0, 'completions/max_length': 244.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 203.0, 'completions/min_terminated_length': 185.0, 'completions/max_terminated_length': 244.0, 'rewards/meter/mean': 0.6535764932632446, 'rewards/meter/std': 0.24456477165222168, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0578637570142746, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/repeat_penalty/mean': 0.49451619386672974, 'rewards/repeat_penalty/std': 0.29470983147621155, 'rewards/near_duplicate_penalty/mean': 0.8792835474014282, 'rewards/near_duplicate_penalty/std': 0.06122273951768875, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.5813867449760437, 'rewards/distinct_2/std': 0.3329528868198395, 'rewards/total_composite/mean': 0.20918208360671997, 'rewards/total_composite/std': 0.10735700279474258, 'reward': 0.20918208360671997, 'reward_std': 0.10735700279474258, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08552273362874985, 'sampling/sampling_logp_difference/max': 1.9656343460083008, 'sampling/importance_sampling_ratio/min': 0.140066996216774, 'sampling/importance_sampling_ratio/mean': 1.0032917261123657, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.37725358456373215, 'clip_ratio/low_mean': 0.05769258085638285, 'clip_ratio/low_min': 0.05769258085638285, 'clip_ratio/high_mean': 0.024382275994867086, 'clip_ratio/high_max': 0.024382275994867086, 'clip_ratio/region_mean': 0.08207485685124993, 'reward_total_mean': 0.20918208360671997, 'reward_meter_mean': 0.6535764932632446, 'reward_meter_std': 0.24456477165222168, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0578637570142746, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.49451619386672974, 'reward_repeat_penalty_std': 0.29470983147621155, 'reward_near_duplicate_penalty_mean': 0.8792835474014282, 'reward_near_duplicate_penalty_std': 0.06122273951768875, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.5813867449760437, 'reward_distinct_2_std': 0.3329528868198395, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.20918208360671997, 'reward_total_composite_std': 0.10735700279474258, 'epoch': 0.03}

 26%|██▌       | 849/3300 [3:22:51<8:27:15, 12.42s/it]INFO 04-12 14:09:03 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 850/3300 [3:23:05<8:40:14, 12.74s/it]2026-04-12 14:09:17,136 | INFO | train_grpo_train | metrics_logged mode=train step=850

                                                      
{'loss': -0.046, 'grad_norm': 3.770644187927246, 'learning_rate': 7.4272727272727275e-06, 'num_tokens': 1729265.0, 'completions/mean_length': 84.25, 'completions/min_length': 22.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 23.142858505249023, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.7202411890029907, 'rewards/meter/std': 0.4407414495944977, '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/lexical_plausibility/mean': 0.9792736172676086, 'rewards/lexical_plausibility/std': 0.03989769518375397, 'rewards/judge_quality/mean': 0.7150000333786011, 'rewards/judge_quality/std': 0.38052597641944885, 'rewards/repeat_penalty/mean': 0.73777174949646, 'rewards/repeat_penalty/std': 0.03458673879504204, 'rewards/near_duplicate_penalty/mean': 0.9836956262588501, 'rewards/near_duplicate_penalty/std': 0.04611567035317421, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.582308828830719, 'rewards/total_composite/std': 0.4445638954639435, 'reward': 0.582308828830719, 'reward_std': 0.4445638954639435, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1147550567984581, 'sampling/sampling_logp_difference/max': 1.2857037782669067, 'sampling/importance_sampling_ratio/min': 0.27645593881607056, 'sampling/importance_sampling_ratio/mean': 1.0257759094238281, 'sampling/importance_sampling_ratio/max': 1.9395263195037842, 'entropy': 0.5739097781479359, 'clip_ratio/low_mean': 0.02223320212215185, 'clip_ratio/low_min': 0.02223320212215185, 'clip_ratio/high_mean': 0.0953409094363451, 'clip_ratio/high_max': 0.0953409094363451, 'clip_ratio/region_mean': 0.11757411155849695, 'reward_total_mean': 0.582308828830719, 'reward_meter_mean': 0.7202411890029907, 'reward_meter_std': 0.4407414495944977, '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_lexical_plausibility_mean': 0.9792736172676086, 'reward_lexical_plausibility_std': 0.03989769518375397, 'reward_repeat_penalty_mean': 0.73777174949646, 'reward_repeat_penalty_std': 0.03458673879504204, 'reward_near_duplicate_penalty_mean': 0.9836956262588501, 'reward_near_duplicate_penalty_std': 0.04611567035317421, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7150000333786011, 'reward_judge_quality_std': 0.38052597641944885, 'reward_total_composite_mean': 0.582308828830719, 'reward_total_composite_std': 0.4445638954639435, 'epoch': 0.03}

 26%|██▌       | 850/3300 [3:23:05<8:40:14, 12.74s/it]INFO 04-12 14:09:17 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


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

 15%|█▌        | 2/13 [00:12<01:06,  6.08s/it]

 23%|██▎       | 3/13 [00:23<01:20,  8.08s/it]

 31%|███       | 4/13 [00:34<01:22,  9.21s/it]

 38%|███▊      | 5/13 [00:45<01:19,  9.99s/it]

 46%|████▌     | 6/13 [00:57<01:13, 10.48s/it]

 54%|█████▍    | 7/13 [01:08<01:04, 10.81s/it]

 62%|██████▏   | 8/13 [01:19<00:54, 10.99s/it]

 69%|██████▉   | 9/13 [01:31<00:44, 11.20s/it]

 77%|███████▋  | 10/13 [01:43<00:34, 11.35s/it]

 85%|████████▍ | 11/13 [01:55<00:23, 11.59s/it]

 92%|█████████▏| 12/13 [02:06<00:11, 11.50s/it]

100%|██████████| 13/13 [02:18<00:00, 11.52s/it]2026-04-12 14:11:48,320 | INFO | train_grpo_train | metrics_logged mode=eval step=850

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 151.1241, 'eval_samples_per_second': 0.688, 'eval_steps_per_second': 0.086, 'eval_num_tokens': 1729265.0, 'eval_completions/mean_length': 249.90384615384616, 'eval_completions/min_length': 51.53846153846154, 'eval_completions/max_length': 511.53846153846155, 'eval_completions/clipped_ratio': 0.22115384615384615, 'eval_completions/mean_terminated_length': 176.33379540076623, 'eval_completions/min_terminated_length': 51.53846153846154, 'eval_completions/max_terminated_length': 365.0, 'eval_rewards/meter/mean': 0.4448819011449814, 'eval_rewards/meter/std': 0.3580018625809596, 'eval_rewards/count_adherence/mean': 0.8912672629723182, 'eval_rewards/count_adherence/std': 0.1746937598173435, 'eval_rewards/arabic_clean/mean': 0.8557692307692307, 'eval_rewards/arabic_clean/std': 0.28682195911040675, 'eval_rewards/lexical_plausibility/mean': 0.9675572973031265, 'eval_rewards/lexical_plausibility/std': 0.05942482429628189, 'eval_rewards/judge_quality/mean': 0.3385576903820038, 'eval_rewards/judge_quality/std': 0.21257938329990095, 'eval_rewards/repeat_penalty/mean': 0.6054709003521845, 'eval_rewards/repeat_penalty/std': 0.37947613573991335, 'eval_rewards/near_duplicate_penalty/mean': 0.8781809302476736, 'eval_rewards/near_duplicate_penalty/std': 0.13536858558654785, 'eval_rewards/opening_diversity/mean': 0.9364497157243582, 'eval_rewards/opening_diversity/std': 0.12932605783526713, 'eval_rewards/distinct_2/mean': 0.671563290632688, 'eval_rewards/distinct_2/std': 0.38163365652927983, 'eval_rewards/total_composite/mean': 0.12671952522717989, 'eval_rewards/total_composite/std': 0.13176435031569922, 'eval_reward': 0.12671952522717989, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.04447327324977288, 'eval_sampling/sampling_logp_difference/max': 1.0257074282719538, 'eval_sampling/importance_sampling_ratio/min': 0.3644731044769287, 'eval_sampling/importance_sampling_ratio/mean': 1.0076959683344915, 'eval_sampling/importance_sampling_ratio/max': 1.5320360935651338, 'eval_entropy': 0.4640323542631589, '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': 0.12671952522717989, 'eval_reward_meter_mean': 0.4448819011449814, 'eval_reward_meter_std': 0.3580018625809596, 'eval_reward_count_adherence_mean': 0.8912672629723182, 'eval_reward_count_adherence_std': 0.1746937598173435, 'eval_reward_arabic_clean_mean': 0.8557692307692307, 'eval_reward_arabic_clean_std': 0.28682195911040675, 'eval_reward_lexical_plausibility_mean': 0.9675572973031265, 'eval_reward_lexical_plausibility_std': 0.05942482429628189, 'eval_reward_repeat_penalty_mean': 0.6054709003521845, 'eval_reward_repeat_penalty_std': 0.37947613573991335, 'eval_reward_near_duplicate_penalty_mean': 0.8781809302476736, 'eval_reward_near_duplicate_penalty_std': 0.13536858558654785, 'eval_reward_opening_diversity_mean': 0.9364497157243582, 'eval_reward_opening_diversity_std': 0.12932605783526713, 'eval_reward_distinct_2_mean': 0.671563290632688, 'eval_reward_distinct_2_std': 0.38163365652927983, 'eval_reward_judge_quality_mean': 0.3385576903820038, 'eval_reward_judge_quality_std': 0.21257938329990095, 'eval_reward_total_composite_mean': 0.12671952522717989, 'eval_reward_total_composite_std': 0.13176435031569922, 'epoch': 0.03}

 26%|██▌       | 850/3300 [3:25:36<8:40:14, 12.74s/it]

100%|██████████| 13/13 [02:18<00:00, 11.52s/it]

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

 26%|██▌       | 851/3300 [3:25:53<40:23:24, 59.37s/it]2026-04-12 14:12:05,319 | INFO | train_grpo_train | metrics_logged mode=train step=851

                                                       
{'loss': -0.1915, 'grad_norm': 3.036870241165161, 'learning_rate': 7.424242424242425e-06, 'num_tokens': 1731652.0, 'completions/mean_length': 238.375, 'completions/min_length': 138.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 147.1666717529297, 'completions/min_terminated_length': 138.0, 'completions/max_terminated_length': 157.0, 'rewards/meter/mean': 0.49326273798942566, 'rewards/meter/std': 0.2799670696258545, 'rewards/count_adherence/mean': 0.9166666269302368, 'rewards/count_adherence/std': 0.17817415297031403, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9327911138534546, 'rewards/lexical_plausibility/std': 0.12464896589517593, 'rewards/judge_quality/mean': 0.2749999761581421, 'rewards/judge_quality/std': 0.16690459847450256, 'rewards/repeat_penalty/mean': 0.8372189998626709, 'rewards/repeat_penalty/std': 0.1295851618051529, 'rewards/near_duplicate_penalty/mean': 0.9472448825836182, 'rewards/near_duplicate_penalty/std': 0.027554435655474663, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8812383413314819, 'rewards/distinct_2/std': 0.11631060391664505, 'rewards/total_composite/mean': 0.15697631239891052, 'rewards/total_composite/std': 0.12170809507369995, 'reward': 0.15697631239891052, 'reward_std': 0.12170809507369995, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13440777361392975, 'sampling/sampling_logp_difference/max': 2.7319352626800537, 'sampling/importance_sampling_ratio/min': 0.06509319692850113, 'sampling/importance_sampling_ratio/mean': 0.996009111404419, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5036326497793198, 'clip_ratio/low_mean': 0.03144668601453304, 'clip_ratio/low_min': 0.03144668601453304, 'clip_ratio/high_mean': 0.078272994607687, 'clip_ratio/high_max': 0.078272994607687, 'clip_ratio/region_mean': 0.10971968062222004, 'reward_total_mean': 0.15697631239891052, 'reward_meter_mean': 0.49326273798942566, 'reward_meter_std': 0.2799670696258545, 'reward_count_adherence_mean': 0.9166666269302368, 'reward_count_adherence_std': 0.17817415297031403, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9327911138534546, 'reward_lexical_plausibility_std': 0.12464896589517593, 'reward_repeat_penalty_mean': 0.8372189998626709, 'reward_repeat_penalty_std': 0.1295851618051529, 'reward_near_duplicate_penalty_mean': 0.9472448825836182, 'reward_near_duplicate_penalty_std': 0.027554435655474663, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8812383413314819, 'reward_distinct_2_std': 0.11631060391664505, 'reward_judge_quality_mean': 0.2749999761581421, 'reward_judge_quality_std': 0.16690459847450256, 'reward_total_composite_mean': 0.15697631239891052, 'reward_total_composite_std': 0.12170809507369995, 'epoch': 0.03}

 26%|██▌       | 851/3300 [3:25:53<40:23:24, 59.37s/it]INFO 04-12 14:12:05 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 852/3300 [3:26:06<31:02:23, 45.65s/it]2026-04-12 14:12:18,934 | INFO | train_grpo_train | metrics_logged mode=train step=852

                                                       
{'loss': 0.0236, 'grad_norm': 5.771389007568359, 'learning_rate': 7.421212121212121e-06, 'num_tokens': 1733330.0, 'completions/mean_length': 117.75, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 61.42857360839844, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 118.0, 'rewards/meter/mean': 0.6570340394973755, 'rewards/meter/std': 0.3390442430973053, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9543074369430542, 'rewards/lexical_plausibility/std': 0.09006527811288834, 'rewards/judge_quality/mean': 0.5587500333786011, 'rewards/judge_quality/std': 0.31669893860816956, 'rewards/repeat_penalty/mean': 0.9256876111030579, 'rewards/repeat_penalty/std': 0.06792078167200089, 'rewards/near_duplicate_penalty/mean': 0.9544930458068848, 'rewards/near_duplicate_penalty/std': 0.04653129726648331, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.969480037689209, 'rewards/distinct_2/std': 0.04637126624584198, 'rewards/total_composite/mean': 0.35104185342788696, 'rewards/total_composite/std': 0.3258068859577179, 'reward': 0.35104185342788696, 'reward_std': 0.3258068859577179, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14230962097644806, 'sampling/sampling_logp_difference/max': 1.5181989669799805, 'sampling/importance_sampling_ratio/min': 0.21910616755485535, 'sampling/importance_sampling_ratio/mean': 1.005537986755371, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8239889964461327, 'clip_ratio/low_mean': 0.02830155473202467, 'clip_ratio/low_min': 0.02830155473202467, 'clip_ratio/high_mean': 0.08166788890957832, 'clip_ratio/high_max': 0.08166788890957832, 'clip_ratio/region_mean': 0.10996944364160299, 'reward_total_mean': 0.35104185342788696, 'reward_meter_mean': 0.6570340394973755, 'reward_meter_std': 0.3390442430973053, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9543074369430542, 'reward_lexical_plausibility_std': 0.09006527811288834, 'reward_repeat_penalty_mean': 0.9256876111030579, 'reward_repeat_penalty_std': 0.06792078167200089, 'reward_near_duplicate_penalty_mean': 0.9544930458068848, 'reward_near_duplicate_penalty_std': 0.04653129726648331, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.969480037689209, 'reward_distinct_2_std': 0.04637126624584198, 'reward_judge_quality_mean': 0.5587500333786011, 'reward_judge_quality_std': 0.31669893860816956, 'reward_total_composite_mean': 0.35104185342788696, 'reward_total_composite_std': 0.3258068859577179, 'epoch': 0.03}

 26%|██▌       | 852/3300 [3:26:06<31:02:23, 45.65s/it]INFO 04-12 14:12:19 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 853/3300 [3:26:21<24:45:31, 36.42s/it]2026-04-12 14:12:33,839 | INFO | train_grpo_train | metrics_logged mode=train step=853

                                                       
{'loss': 0.0951, 'grad_norm': 1.5703561305999756, 'learning_rate': 7.4181818181818185e-06, 'num_tokens': 1737230.0, 'completions/mean_length': 457.5, 'completions/min_length': 389.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 424.8000183105469, 'completions/min_terminated_length': 389.0, 'completions/max_terminated_length': 492.0, 'rewards/meter/mean': 0.6266857385635376, 'rewards/meter/std': 0.362479567527771, 'rewards/count_adherence/mean': 0.6538461446762085, 'rewards/count_adherence/std': 0.23619985580444336, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9819031357765198, 'rewards/lexical_plausibility/std': 0.04779762774705887, 'rewards/judge_quality/mean': 0.13750000298023224, 'rewards/judge_quality/std': 0.0353553406894207, 'rewards/repeat_penalty/mean': 0.12918885052204132, 'rewards/repeat_penalty/std': 0.34368520975112915, 'rewards/near_duplicate_penalty/mean': 0.7267413139343262, 'rewards/near_duplicate_penalty/std': 0.153737410902977, 'rewards/opening_diversity/mean': 0.7754915952682495, 'rewards/opening_diversity/std': 0.3050152659416199, 'rewards/distinct_2/mean': 0.13968680799007416, 'rewards/distinct_2/std': 0.348369836807251, 'rewards/total_composite/mean': 0.05157095566391945, 'rewards/total_composite/std': 0.037722326815128326, 'reward': 0.05157095566391945, 'reward_std': 0.03772232308983803, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05474357679486275, 'sampling/sampling_logp_difference/max': 2.797646999359131, 'sampling/importance_sampling_ratio/min': 0.06095331534743309, 'sampling/importance_sampling_ratio/mean': 1.0069084167480469, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1865599974989891, 'clip_ratio/low_mean': 0.018432441633194685, 'clip_ratio/low_min': 0.018432441633194685, 'clip_ratio/high_mean': 0.011336264666169882, 'clip_ratio/high_max': 0.011336264666169882, 'clip_ratio/region_mean': 0.029768706299364567, 'reward_total_mean': 0.05157095566391945, 'reward_meter_mean': 0.6266857385635376, 'reward_meter_std': 0.362479567527771, 'reward_count_adherence_mean': 0.6538461446762085, 'reward_count_adherence_std': 0.23619985580444336, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9819031357765198, 'reward_lexical_plausibility_std': 0.04779762774705887, 'reward_repeat_penalty_mean': 0.12918885052204132, 'reward_repeat_penalty_std': 0.34368520975112915, 'reward_near_duplicate_penalty_mean': 0.7267413139343262, 'reward_near_duplicate_penalty_std': 0.153737410902977, 'reward_opening_diversity_mean': 0.7754915952682495, 'reward_opening_diversity_std': 0.3050152659416199, 'reward_distinct_2_mean': 0.13968680799007416, 'reward_distinct_2_std': 0.348369836807251, 'reward_judge_quality_mean': 0.13750000298023224, 'reward_judge_quality_std': 0.0353553406894207, 'reward_total_composite_mean': 0.05157095566391945, 'reward_total_composite_std': 0.037722326815128326, 'epoch': 0.03}

 26%|██▌       | 853/3300 [3:26:21<24:45:31, 36.42s/it]INFO 04-12 14:12:34 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 854/3300 [3:26:34<19:59:17, 29.42s/it]2026-04-12 14:12:46,910 | INFO | train_grpo_train | metrics_logged mode=train step=854

                                                       
{'loss': -0.1222, 'grad_norm': 3.162191152572632, 'learning_rate': 7.415151515151515e-06, 'num_tokens': 1739061.0, 'completions/mean_length': 186.875, 'completions/min_length': 66.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 78.5, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 85.0, 'rewards/meter/mean': 0.35487979650497437, 'rewards/meter/std': 0.17712539434432983, '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/lexical_plausibility/mean': 0.93937087059021, 'rewards/lexical_plausibility/std': 0.11311454325914383, 'rewards/judge_quality/mean': 0.5450000166893005, 'rewards/judge_quality/std': 0.3766392767429352, 'rewards/repeat_penalty/mean': 0.9895455241203308, 'rewards/repeat_penalty/std': 0.008139162324368954, 'rewards/near_duplicate_penalty/mean': 0.9900454878807068, 'rewards/near_duplicate_penalty/std': 0.008581546135246754, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9994992017745972, 'rewards/distinct_2/std': 0.0014164713211357594, 'rewards/total_composite/mean': 0.22739753127098083, 'rewards/total_composite/std': 0.17319437861442566, 'reward': 0.22739753127098083, 'reward_std': 0.17319437861442566, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1522524654865265, 'sampling/sampling_logp_difference/max': 1.7845678329467773, 'sampling/importance_sampling_ratio/min': 0.16786959767341614, 'sampling/importance_sampling_ratio/mean': 0.9787825345993042, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.42490098997950554, 'clip_ratio/low_mean': 0.01617647148668766, 'clip_ratio/low_min': 0.01617647148668766, 'clip_ratio/high_mean': 0.0963799711316824, 'clip_ratio/high_max': 0.0963799711316824, 'clip_ratio/region_mean': 0.11255644261837006, 'reward_total_mean': 0.22739753127098083, 'reward_meter_mean': 0.35487979650497437, 'reward_meter_std': 0.17712539434432983, '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_lexical_plausibility_mean': 0.93937087059021, 'reward_lexical_plausibility_std': 0.11311454325914383, 'reward_repeat_penalty_mean': 0.9895455241203308, 'reward_repeat_penalty_std': 0.008139162324368954, 'reward_near_duplicate_penalty_mean': 0.9900454878807068, 'reward_near_duplicate_penalty_std': 0.008581546135246754, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9994992017745972, 'reward_distinct_2_std': 0.0014164713211357594, 'reward_judge_quality_mean': 0.5450000166893005, 'reward_judge_quality_std': 0.3766392767429352, 'reward_total_composite_mean': 0.22739753127098083, 'reward_total_composite_std': 0.17319437861442566, 'epoch': 0.03}

 26%|██▌       | 854/3300 [3:26:34<19:59:17, 29.42s/it]INFO 04-12 14:12:47 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 855/3300 [3:26:47<16:36:23, 24.45s/it]2026-04-12 14:12:59,776 | INFO | train_grpo_train | metrics_logged mode=train step=855

                                                       
{'loss': -0.0805, 'grad_norm': 5.049601078033447, 'learning_rate': 7.412121212121213e-06, 'num_tokens': 1740582.0, 'completions/mean_length': 102.125, 'completions/min_length': 38.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 43.57143020629883, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.9404455423355103, 'rewards/meter/std': 0.10320204496383667, '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/lexical_plausibility/mean': 0.9707344174385071, 'rewards/lexical_plausibility/std': 0.0827755481004715, 'rewards/judge_quality/mean': 0.6349999904632568, 'rewards/judge_quality/std': 0.3443005383014679, 'rewards/repeat_penalty/mean': 0.9731278419494629, 'rewards/repeat_penalty/std': 0.026308590546250343, 'rewards/near_duplicate_penalty/mean': 0.9731278419494629, 'rewards/near_duplicate_penalty/std': 0.026308590546250343, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6001728773117065, 'rewards/total_composite/std': 0.3590646982192993, 'reward': 0.6001728773117065, 'reward_std': 0.35906466841697693, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10709303617477417, 'sampling/sampling_logp_difference/max': 1.1990714073181152, 'sampling/importance_sampling_ratio/min': 0.30147403478622437, 'sampling/importance_sampling_ratio/mean': 1.0106830596923828, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4722936786711216, 'clip_ratio/low_mean': 0.019832042045891285, 'clip_ratio/low_min': 0.019832042045891285, 'clip_ratio/high_mean': 0.07080832682549953, 'clip_ratio/high_max': 0.07080832682549953, 'clip_ratio/region_mean': 0.09064036887139082, 'reward_total_mean': 0.6001728773117065, 'reward_meter_mean': 0.9404455423355103, 'reward_meter_std': 0.10320204496383667, '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_lexical_plausibility_mean': 0.9707344174385071, 'reward_lexical_plausibility_std': 0.0827755481004715, 'reward_repeat_penalty_mean': 0.9731278419494629, 'reward_repeat_penalty_std': 0.026308590546250343, 'reward_near_duplicate_penalty_mean': 0.9731278419494629, 'reward_near_duplicate_penalty_std': 0.026308590546250343, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6349999904632568, 'reward_judge_quality_std': 0.3443005383014679, 'reward_total_composite_mean': 0.6001728773117065, 'reward_total_composite_std': 0.3590646982192993, 'epoch': 0.03}

 26%|██▌       | 855/3300 [3:26:47<16:36:23, 24.45s/it]INFO 04-12 14:13:00 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 856/3300 [3:27:00<14:16:56, 21.04s/it]2026-04-12 14:13:12,844 | INFO | train_grpo_train | metrics_logged mode=train step=856

                                                       
{'loss': -0.0728, 'grad_norm': 1.8549288511276245, 'learning_rate': 7.40909090909091e-06, 'num_tokens': 1742168.0, 'completions/mean_length': 220.25, 'completions/min_length': 42.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 45.20000076293945, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.7261562347412109, 'rewards/meter/std': 0.38776305317878723, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.915650486946106, 'rewards/lexical_plausibility/std': 0.11713753640651703, 'rewards/judge_quality/mean': 0.3462499976158142, 'rewards/judge_quality/std': 0.29875636100769043, 'rewards/repeat_penalty/mean': 0.952872633934021, 'rewards/repeat_penalty/std': 0.06170550361275673, 'rewards/near_duplicate_penalty/mean': 0.9707337021827698, 'rewards/near_duplicate_penalty/std': 0.023901375010609627, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9810642600059509, 'rewards/distinct_2/std': 0.04894232377409935, 'rewards/total_composite/mean': 0.30604180693626404, 'rewards/total_composite/std': 0.3092731237411499, 'reward': 0.30604180693626404, 'reward_std': 0.3092730939388275, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1227310299873352, 'sampling/sampling_logp_difference/max': 1.4008054733276367, 'sampling/importance_sampling_ratio/min': 0.2463984340429306, 'sampling/importance_sampling_ratio/mean': 0.9991787672042847, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4100022315979004, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.08825570717453957, 'clip_ratio/high_max': 0.08825570717453957, 'clip_ratio/region_mean': 0.08825570717453957, 'reward_total_mean': 0.30604180693626404, 'reward_meter_mean': 0.7261562347412109, 'reward_meter_std': 0.38776305317878723, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.915650486946106, 'reward_lexical_plausibility_std': 0.11713753640651703, 'reward_repeat_penalty_mean': 0.952872633934021, 'reward_repeat_penalty_std': 0.06170550361275673, 'reward_near_duplicate_penalty_mean': 0.9707337021827698, 'reward_near_duplicate_penalty_std': 0.023901375010609627, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9810642600059509, 'reward_distinct_2_std': 0.04894232377409935, 'reward_judge_quality_mean': 0.3462499976158142, 'reward_judge_quality_std': 0.29875636100769043, 'reward_total_composite_mean': 0.30604180693626404, 'reward_total_composite_std': 0.3092731237411499, 'epoch': 0.03}

 26%|██▌       | 856/3300 [3:27:00<14:16:56, 21.04s/it]INFO 04-12 14:13:13 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 857/3300 [3:27:14<12:43:20, 18.75s/it]2026-04-12 14:13:26,260 | INFO | train_grpo_train | metrics_logged mode=train step=857

                                                       
{'loss': -0.0541, 'grad_norm': 2.7563040256500244, 'learning_rate': 7.406060606060607e-06, 'num_tokens': 1743617.0, 'completions/mean_length': 223.125, 'completions/min_length': 44.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 49.79999923706055, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.27811571955680847, 'rewards/meter/std': 0.3214670419692993, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.8602814674377441, 'rewards/lexical_plausibility/std': 0.15611913800239563, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.22320714592933655, 'rewards/repeat_penalty/mean': 0.9917591214179993, 'rewards/repeat_penalty/std': 0.01286725327372551, 'rewards/near_duplicate_penalty/mean': 0.9917591214179993, 'rewards/near_duplicate_penalty/std': 0.01286725327372551, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.12017762660980225, 'rewards/total_composite/std': 0.17518144845962524, 'reward': 0.12017762660980225, 'reward_std': 0.17518143355846405, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1400035172700882, 'sampling/sampling_logp_difference/max': 1.4476470947265625, 'sampling/importance_sampling_ratio/min': 0.23512287437915802, 'sampling/importance_sampling_ratio/mean': 1.0204826593399048, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5590995326638222, 'clip_ratio/low_mean': 0.028879476711153984, 'clip_ratio/low_min': 0.028879476711153984, 'clip_ratio/high_mean': 0.03676243592053652, 'clip_ratio/high_max': 0.03676243592053652, 'clip_ratio/region_mean': 0.0656419126316905, 'reward_total_mean': 0.12017762660980225, 'reward_meter_mean': 0.27811571955680847, 'reward_meter_std': 0.3214670419692993, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.8602814674377441, 'reward_lexical_plausibility_std': 0.15611913800239563, 'reward_repeat_penalty_mean': 0.9917591214179993, 'reward_repeat_penalty_std': 0.01286725327372551, 'reward_near_duplicate_penalty_mean': 0.9917591214179993, 'reward_near_duplicate_penalty_std': 0.01286725327372551, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.22320714592933655, 'reward_total_composite_mean': 0.12017762660980225, 'reward_total_composite_std': 0.17518144845962524, 'epoch': 0.03}

 26%|██▌       | 857/3300 [3:27:14<12:43:20, 18.75s/it]INFO 04-12 14:13:26 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 858/3300 [3:27:27<11:34:43, 17.07s/it]2026-04-12 14:13:39,403 | INFO | train_grpo_train | metrics_logged mode=train step=858

                                                       
{'loss': -0.0145, 'grad_norm': 3.7855429649353027, 'learning_rate': 7.403030303030304e-06, 'num_tokens': 1746891.0, 'completions/mean_length': 292.25, 'completions/min_length': 192.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 260.8571472167969, 'completions/min_terminated_length': 192.0, 'completions/max_terminated_length': 378.0, 'rewards/meter/mean': 0.33917608857154846, 'rewards/meter/std': 0.2654321789741516, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.21128857135772705, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.974175751209259, 'rewards/lexical_plausibility/std': 0.07304198294878006, 'rewards/judge_quality/mean': 0.20000000298023224, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.34017446637153625, 'rewards/repeat_penalty/std': 0.19788828492164612, 'rewards/near_duplicate_penalty/mean': 0.8011940717697144, 'rewards/near_duplicate_penalty/std': 0.10140984505414963, 'rewards/opening_diversity/mean': 0.9831730723381042, 'rewards/opening_diversity/std': 0.04759372025728226, 'rewards/distinct_2/mean': 0.4517394006252289, 'rewards/distinct_2/std': 0.1864631175994873, 'rewards/total_composite/mean': 0.05384580045938492, 'rewards/total_composite/std': 0.04389374330639839, 'reward': 0.05384580045938492, 'reward_std': 0.04389374330639839, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09047599136829376, 'sampling/sampling_logp_difference/max': 3.042069911956787, 'sampling/importance_sampling_ratio/min': 0.04773597791790962, 'sampling/importance_sampling_ratio/mean': 1.0003856420516968, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3925344701856375, 'clip_ratio/low_mean': 0.03446326730772853, 'clip_ratio/low_min': 0.03446326730772853, 'clip_ratio/high_mean': 0.03539368836209178, 'clip_ratio/high_max': 0.03539368836209178, 'clip_ratio/region_mean': 0.06985695566982031, 'reward_total_mean': 0.05384580045938492, 'reward_meter_mean': 0.33917608857154846, 'reward_meter_std': 0.2654321789741516, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.21128857135772705, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.974175751209259, 'reward_lexical_plausibility_std': 0.07304198294878006, 'reward_repeat_penalty_mean': 0.34017446637153625, 'reward_repeat_penalty_std': 0.19788828492164612, 'reward_near_duplicate_penalty_mean': 0.8011940717697144, 'reward_near_duplicate_penalty_std': 0.10140984505414963, 'reward_opening_diversity_mean': 0.9831730723381042, 'reward_opening_diversity_std': 0.04759372025728226, 'reward_distinct_2_mean': 0.4517394006252289, 'reward_distinct_2_std': 0.1864631175994873, 'reward_judge_quality_mean': 0.20000000298023224, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.05384580045938492, 'reward_total_composite_std': 0.04389374330639839, 'epoch': 0.03}

 26%|██▌       | 858/3300 [3:27:27<11:34:43, 17.07s/it]INFO 04-12 14:13:39 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 859/3300 [3:27:41<10:55:20, 16.11s/it]2026-04-12 14:13:53,271 | INFO | train_grpo_train | metrics_logged mode=train step=859

                                                       
{'loss': -0.0723, 'grad_norm': 4.451854228973389, 'learning_rate': 7.4e-06, 'num_tokens': 1749133.0, 'completions/mean_length': 159.25, 'completions/min_length': 102.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 108.85714721679688, 'completions/min_terminated_length': 102.0, 'completions/max_terminated_length': 119.0, 'rewards/meter/mean': 0.3008106052875519, 'rewards/meter/std': 0.24272847175598145, '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/lexical_plausibility/mean': 0.972153902053833, 'rewards/lexical_plausibility/std': 0.0459597073495388, 'rewards/judge_quality/mean': 0.5299999713897705, 'rewards/judge_quality/std': 0.346533864736557, 'rewards/repeat_penalty/mean': 0.9406341314315796, 'rewards/repeat_penalty/std': 0.06208498775959015, 'rewards/near_duplicate_penalty/mean': 0.9765318632125854, 'rewards/near_duplicate_penalty/std': 0.01911858841776848, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9630399942398071, 'rewards/distinct_2/std': 0.057189639657735825, 'rewards/total_composite/mean': 0.17408716678619385, 'rewards/total_composite/std': 0.1432017832994461, 'reward': 0.17408716678619385, 'reward_std': 0.1432017832994461, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11997110396623611, 'sampling/sampling_logp_difference/max': 2.8195769786834717, 'sampling/importance_sampling_ratio/min': 0.05963116139173508, 'sampling/importance_sampling_ratio/mean': 0.9923701286315918, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4308541715145111, 'clip_ratio/low_mean': 0.04663661774247885, 'clip_ratio/low_min': 0.04663661774247885, 'clip_ratio/high_mean': 0.06251640524715185, 'clip_ratio/high_max': 0.06251640524715185, 'clip_ratio/region_mean': 0.1091530229896307, 'reward_total_mean': 0.17408716678619385, 'reward_meter_mean': 0.3008106052875519, 'reward_meter_std': 0.24272847175598145, '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_lexical_plausibility_mean': 0.972153902053833, 'reward_lexical_plausibility_std': 0.0459597073495388, 'reward_repeat_penalty_mean': 0.9406341314315796, 'reward_repeat_penalty_std': 0.06208498775959015, 'reward_near_duplicate_penalty_mean': 0.9765318632125854, 'reward_near_duplicate_penalty_std': 0.01911858841776848, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9630399942398071, 'reward_distinct_2_std': 0.057189639657735825, 'reward_judge_quality_mean': 0.5299999713897705, 'reward_judge_quality_std': 0.346533864736557, 'reward_total_composite_mean': 0.17408716678619385, 'reward_total_composite_std': 0.1432017832994461, 'epoch': 0.03}

 26%|██▌       | 859/3300 [3:27:41<10:55:20, 16.11s/it]INFO 04-12 14:13:53 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 860/3300 [3:27:50<9:30:33, 14.03s/it] 2026-04-12 14:14:02,453 | INFO | train_grpo_train | metrics_logged mode=train step=860

                                                      
{'loss': 0.0024, 'grad_norm': 7.0358099937438965, 'learning_rate': 7.396969696969698e-06, 'num_tokens': 1751326.0, 'completions/mean_length': 91.125, 'completions/min_length': 81.0, 'completions/max_length': 114.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 91.125, 'completions/min_terminated_length': 81.0, 'completions/max_terminated_length': 114.0, 'rewards/meter/mean': 0.7421751022338867, 'rewards/meter/std': 0.3366747796535492, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5174999833106995, 'rewards/judge_quality/std': 0.2522329092025757, 'rewards/repeat_penalty/mean': 0.9613125324249268, 'rewards/repeat_penalty/std': 0.049800075590610504, 'rewards/near_duplicate_penalty/mean': 0.9764053225517273, 'rewards/near_duplicate_penalty/std': 0.020267708227038383, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9839743375778198, 'rewards/distinct_2/std': 0.031795889139175415, 'rewards/total_composite/mean': 0.40260064601898193, 'rewards/total_composite/std': 0.31393536925315857, 'reward': 0.40260064601898193, 'reward_std': 0.31393536925315857, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1001368835568428, 'sampling/sampling_logp_difference/max': 1.8461298942565918, 'sampling/importance_sampling_ratio/min': 0.15784688293933868, 'sampling/importance_sampling_ratio/mean': 0.9935073852539062, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4307049550116062, 'clip_ratio/low_mean': 0.07043880689889193, 'clip_ratio/low_min': 0.07043880689889193, 'clip_ratio/high_mean': 0.03290719632059336, 'clip_ratio/high_max': 0.03290719632059336, 'clip_ratio/region_mean': 0.10334600321948528, 'reward_total_mean': 0.40260064601898193, 'reward_meter_mean': 0.7421751022338867, 'reward_meter_std': 0.3366747796535492, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9613125324249268, 'reward_repeat_penalty_std': 0.049800075590610504, 'reward_near_duplicate_penalty_mean': 0.9764053225517273, 'reward_near_duplicate_penalty_std': 0.020267708227038383, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9839743375778198, 'reward_distinct_2_std': 0.031795889139175415, 'reward_judge_quality_mean': 0.5174999833106995, 'reward_judge_quality_std': 0.2522329092025757, 'reward_total_composite_mean': 0.40260064601898193, 'reward_total_composite_std': 0.31393536925315857, 'epoch': 0.03}

 26%|██▌       | 860/3300 [3:27:50<9:30:33, 14.03s/it]INFO 04-12 14:14:02 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 861/3300 [3:28:03<9:16:56, 13.70s/it]2026-04-12 14:14:15,385 | INFO | train_grpo_train | metrics_logged mode=train step=861

                                                      
{'loss': -0.0756, 'grad_norm': 6.31519079208374, 'learning_rate': 7.393939393939395e-06, 'num_tokens': 1753433.0, 'completions/mean_length': 147.375, 'completions/min_length': 88.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 95.28572082519531, 'completions/min_terminated_length': 88.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.48967862129211426, 'rewards/meter/std': 0.3885529637336731, '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/lexical_plausibility/mean': 0.9755135774612427, 'rewards/lexical_plausibility/std': 0.06925806403160095, 'rewards/judge_quality/mean': 0.5137500166893005, 'rewards/judge_quality/std': 0.3586059510707855, 'rewards/repeat_penalty/mean': 0.5996940732002258, 'rewards/repeat_penalty/std': 0.24790264666080475, 'rewards/near_duplicate_penalty/mean': 0.8497533798217773, 'rewards/near_duplicate_penalty/std': 0.08299332112073898, 'rewards/opening_diversity/mean': 0.9765625, 'rewards/opening_diversity/std': 0.032346826046705246, 'rewards/distinct_2/mean': 0.7284420728683472, 'rewards/distinct_2/std': 0.2116343379020691, 'rewards/total_composite/mean': 0.2513570785522461, 'rewards/total_composite/std': 0.28744709491729736, 'reward': 0.2513570785522461, 'reward_std': 0.28744709491729736, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08761265128850937, 'sampling/sampling_logp_difference/max': 2.0567173957824707, 'sampling/importance_sampling_ratio/min': 0.12787304818630219, 'sampling/importance_sampling_ratio/mean': 1.0038927793502808, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3343055620789528, 'clip_ratio/low_mean': 0.034244128968566656, 'clip_ratio/low_min': 0.034244128968566656, 'clip_ratio/high_mean': 0.042165616527199745, 'clip_ratio/high_max': 0.042165616527199745, 'clip_ratio/region_mean': 0.0764097454957664, 'reward_total_mean': 0.2513570785522461, 'reward_meter_mean': 0.48967862129211426, 'reward_meter_std': 0.3885529637336731, '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_lexical_plausibility_mean': 0.9755135774612427, 'reward_lexical_plausibility_std': 0.06925806403160095, 'reward_repeat_penalty_mean': 0.5996940732002258, 'reward_repeat_penalty_std': 0.24790264666080475, 'reward_near_duplicate_penalty_mean': 0.8497533798217773, 'reward_near_duplicate_penalty_std': 0.08299332112073898, 'reward_opening_diversity_mean': 0.9765625, 'reward_opening_diversity_std': 0.032346826046705246, 'reward_distinct_2_mean': 0.7284420728683472, 'reward_distinct_2_std': 0.2116343379020691, 'reward_judge_quality_mean': 0.5137500166893005, 'reward_judge_quality_std': 0.3586059510707855, 'reward_total_composite_mean': 0.2513570785522461, 'reward_total_composite_std': 0.28744709491729736, 'epoch': 0.03}

 26%|██▌       | 861/3300 [3:28:03<9:16:56, 13.70s/it]INFO 04-12 14:14:15 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 862/3300 [3:28:16<9:05:01, 13.41s/it]2026-04-12 14:14:28,122 | INFO | train_grpo_train | metrics_logged mode=train step=862

                                                      
{'loss': -0.1084, 'grad_norm': 4.4236555099487305, 'learning_rate': 7.390909090909092e-06, 'num_tokens': 1754964.0, 'completions/mean_length': 126.375, 'completions/min_length': 57.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 71.28572082519531, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.7048521041870117, 'rewards/meter/std': 0.2765464186668396, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.2651650309562683, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9603685140609741, 'rewards/lexical_plausibility/std': 0.10133814066648483, 'rewards/judge_quality/mean': 0.13750000298023224, 'rewards/judge_quality/std': 0.0353553406894207, 'rewards/repeat_penalty/mean': 0.5826496481895447, 'rewards/repeat_penalty/std': 0.24502216279506683, 'rewards/near_duplicate_penalty/mean': 0.8624882102012634, 'rewards/near_duplicate_penalty/std': 0.10935547202825546, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.6983522176742554, 'rewards/distinct_2/std': 0.2373177409172058, 'rewards/total_composite/mean': 0.09684468805789948, 'rewards/total_composite/std': 0.055279623717069626, 'reward': 0.09684468805789948, 'reward_std': 0.055279623717069626, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12048140913248062, 'sampling/sampling_logp_difference/max': 1.75498366355896, 'sampling/importance_sampling_ratio/min': 0.20727141201496124, 'sampling/importance_sampling_ratio/mean': 1.0190989971160889, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6958971619606018, 'clip_ratio/low_mean': 0.03912902623414993, 'clip_ratio/low_min': 0.03912902623414993, 'clip_ratio/high_mean': 0.07419338077306747, 'clip_ratio/high_max': 0.07419338077306747, 'clip_ratio/region_mean': 0.11332240700721741, 'reward_total_mean': 0.09684468805789948, 'reward_meter_mean': 0.7048521041870117, 'reward_meter_std': 0.2765464186668396, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.2651650309562683, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9603685140609741, 'reward_lexical_plausibility_std': 0.10133814066648483, 'reward_repeat_penalty_mean': 0.5826496481895447, 'reward_repeat_penalty_std': 0.24502216279506683, 'reward_near_duplicate_penalty_mean': 0.8624882102012634, 'reward_near_duplicate_penalty_std': 0.10935547202825546, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.6983522176742554, 'reward_distinct_2_std': 0.2373177409172058, 'reward_judge_quality_mean': 0.13750000298023224, 'reward_judge_quality_std': 0.0353553406894207, 'reward_total_composite_mean': 0.09684468805789948, 'reward_total_composite_std': 0.055279623717069626, 'epoch': 0.03}

 26%|██▌       | 862/3300 [3:28:16<9:05:01, 13.41s/it]INFO 04-12 14:14:28 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 863/3300 [3:28:30<9:11:33, 13.58s/it]2026-04-12 14:14:42,092 | INFO | train_grpo_train | metrics_logged mode=train step=863

                                                      
{'loss': -0.1476, 'grad_norm': 3.4566609859466553, 'learning_rate': 7.3878787878787885e-06, 'num_tokens': 1757296.0, 'completions/mean_length': 170.5, 'completions/min_length': 117.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 121.71429443359375, 'completions/min_terminated_length': 117.0, 'completions/max_terminated_length': 131.0, 'rewards/meter/mean': 0.7879217863082886, 'rewards/meter/std': 0.3524387776851654, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.956170916557312, 'rewards/lexical_plausibility/std': 0.12396745383739471, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.14078859984874725, 'rewards/repeat_penalty/mean': 0.5562999248504639, 'rewards/repeat_penalty/std': 0.295958936214447, 'rewards/near_duplicate_penalty/mean': 0.8740394115447998, 'rewards/near_duplicate_penalty/std': 0.08502187579870224, 'rewards/opening_diversity/mean': 0.9750000238418579, 'rewards/opening_diversity/std': 0.04629101976752281, 'rewards/distinct_2/mean': 0.6530981063842773, 'rewards/distinct_2/std': 0.25794315338134766, 'rewards/total_composite/mean': 0.27703890204429626, 'rewards/total_composite/std': 0.15440022945404053, 'reward': 0.27703890204429626, 'reward_std': 0.15440022945404053, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10070232301950455, 'sampling/sampling_logp_difference/max': 1.567529320716858, 'sampling/importance_sampling_ratio/min': 0.2085598260164261, 'sampling/importance_sampling_ratio/mean': 1.0005910396575928, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5294371768832207, 'clip_ratio/low_mean': 0.014707273803651333, 'clip_ratio/low_min': 0.014707273803651333, 'clip_ratio/high_mean': 0.07067823689430952, 'clip_ratio/high_max': 0.07067823689430952, 'clip_ratio/region_mean': 0.08538551069796085, 'reward_total_mean': 0.27703890204429626, 'reward_meter_mean': 0.7879217863082886, 'reward_meter_std': 0.3524387776851654, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.956170916557312, 'reward_lexical_plausibility_std': 0.12396745383739471, 'reward_repeat_penalty_mean': 0.5562999248504639, 'reward_repeat_penalty_std': 0.295958936214447, 'reward_near_duplicate_penalty_mean': 0.8740394115447998, 'reward_near_duplicate_penalty_std': 0.08502187579870224, 'reward_opening_diversity_mean': 0.9750000238418579, 'reward_opening_diversity_std': 0.04629101976752281, 'reward_distinct_2_mean': 0.6530981063842773, 'reward_distinct_2_std': 0.25794315338134766, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.14078859984874725, 'reward_total_composite_mean': 0.27703890204429626, 'reward_total_composite_std': 0.15440022945404053, 'epoch': 0.03}

 26%|██▌       | 863/3300 [3:28:30<9:11:33, 13.58s/it]INFO 04-12 14:14:42 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 864/3300 [3:28:43<9:07:34, 13.49s/it]2026-04-12 14:14:55,375 | INFO | train_grpo_train | metrics_logged mode=train step=864

                                                      
{'loss': -0.0882, 'grad_norm': 1.6613599061965942, 'learning_rate': 7.384848484848486e-06, 'num_tokens': 1758751.0, 'completions/mean_length': 224.875, 'completions/min_length': 50.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 52.60000228881836, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.6435930728912354, 'rewards/meter/std': 0.3932993412017822, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.902877926826477, 'rewards/lexical_plausibility/std': 0.10823936015367508, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.1927248239517212, 'rewards/repeat_penalty/mean': 0.5592107772827148, 'rewards/repeat_penalty/std': 0.32434093952178955, 'rewards/near_duplicate_penalty/mean': 0.8366586565971375, 'rewards/near_duplicate_penalty/std': 0.1928642988204956, 'rewards/opening_diversity/mean': 0.71875, 'rewards/opening_diversity/std': 0.23857837915420532, 'rewards/distinct_2/mean': 0.8448572158813477, 'rewards/distinct_2/std': 0.2168506532907486, 'rewards/total_composite/mean': 0.20020878314971924, 'rewards/total_composite/std': 0.18904922902584076, 'reward': 0.20020878314971924, 'reward_std': 0.18904922902584076, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13655053079128265, 'sampling/sampling_logp_difference/max': 2.8553378582000732, 'sampling/importance_sampling_ratio/min': 0.057536378502845764, 'sampling/importance_sampling_ratio/mean': 0.9979873299598694, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3610412999987602, 'clip_ratio/low_mean': 0.014999999664723873, 'clip_ratio/low_min': 0.014999999664723873, 'clip_ratio/high_mean': 0.05459588207304478, 'clip_ratio/high_max': 0.05459588207304478, 'clip_ratio/region_mean': 0.06959588173776865, 'reward_total_mean': 0.20020878314971924, 'reward_meter_mean': 0.6435930728912354, 'reward_meter_std': 0.3932993412017822, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.902877926826477, 'reward_lexical_plausibility_std': 0.10823936015367508, 'reward_repeat_penalty_mean': 0.5592107772827148, 'reward_repeat_penalty_std': 0.32434093952178955, 'reward_near_duplicate_penalty_mean': 0.8366586565971375, 'reward_near_duplicate_penalty_std': 0.1928642988204956, 'reward_opening_diversity_mean': 0.71875, 'reward_opening_diversity_std': 0.23857837915420532, 'reward_distinct_2_mean': 0.8448572158813477, 'reward_distinct_2_std': 0.2168506532907486, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.1927248239517212, 'reward_total_composite_mean': 0.20020878314971924, 'reward_total_composite_std': 0.18904922902584076, 'epoch': 0.03}

 26%|██▌       | 864/3300 [3:28:43<9:07:34, 13.49s/it]INFO 04-12 14:14:55 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 865/3300 [3:28:52<8:10:49, 12.09s/it]2026-04-12 14:15:04,217 | INFO | train_grpo_train | metrics_logged mode=train step=865

                                                      
{'loss': 0.0598, 'grad_norm': 10.180727005004883, 'learning_rate': 7.381818181818182e-06, 'num_tokens': 1760494.0, 'completions/mean_length': 60.875, 'completions/min_length': 55.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.875, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.7840351462364197, 'rewards/meter/std': 0.3521861135959625, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4337500035762787, 'rewards/judge_quality/std': 0.21999594569206238, 'rewards/repeat_penalty/mean': 0.5899261832237244, 'rewards/repeat_penalty/std': 0.2448299080133438, 'rewards/near_duplicate_penalty/mean': 0.8201501965522766, 'rewards/near_duplicate_penalty/std': 0.13156695663928986, 'rewards/opening_diversity/mean': 0.875, 'rewards/opening_diversity/std': 0.13363061845302582, 'rewards/distinct_2/mean': 0.7801418304443359, 'rewards/distinct_2/std': 0.14881575107574463, 'rewards/total_composite/mean': 0.28938913345336914, 'rewards/total_composite/std': 0.12707412242889404, 'reward': 0.28938913345336914, 'reward_std': 0.12707412242889404, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11234337836503983, 'sampling/sampling_logp_difference/max': 1.9080142974853516, 'sampling/importance_sampling_ratio/min': 0.14837472140789032, 'sampling/importance_sampling_ratio/mean': 1.0136069059371948, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5191529989242554, 'clip_ratio/low_mean': 0.030281012412160635, 'clip_ratio/low_min': 0.030281012412160635, 'clip_ratio/high_mean': 0.0663289362564683, 'clip_ratio/high_max': 0.0663289362564683, 'clip_ratio/region_mean': 0.09660994866862893, 'reward_total_mean': 0.28938913345336914, 'reward_meter_mean': 0.7840351462364197, 'reward_meter_std': 0.3521861135959625, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.5899261832237244, 'reward_repeat_penalty_std': 0.2448299080133438, 'reward_near_duplicate_penalty_mean': 0.8201501965522766, 'reward_near_duplicate_penalty_std': 0.13156695663928986, 'reward_opening_diversity_mean': 0.875, 'reward_opening_diversity_std': 0.13363061845302582, 'reward_distinct_2_mean': 0.7801418304443359, 'reward_distinct_2_std': 0.14881575107574463, 'reward_judge_quality_mean': 0.4337500035762787, 'reward_judge_quality_std': 0.21999594569206238, 'reward_total_composite_mean': 0.28938913345336914, 'reward_total_composite_std': 0.12707412242889404, 'epoch': 0.03}

 26%|██▌       | 865/3300 [3:28:52<8:10:49, 12.09s/it]INFO 04-12 14:15:04 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 866/3300 [3:29:05<8:23:02, 12.40s/it]2026-04-12 14:15:17,328 | INFO | train_grpo_train | metrics_logged mode=train step=866

                                                      
{'loss': -0.094, 'grad_norm': 3.834961175918579, 'learning_rate': 7.378787878787879e-06, 'num_tokens': 1762115.0, 'completions/mean_length': 108.625, 'completions/min_length': 46.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 51.000003814697266, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.7777613401412964, 'rewards/meter/std': 0.34793820977211, '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/lexical_plausibility/mean': 0.9689165353775024, 'rewards/lexical_plausibility/std': 0.0879172757267952, 'rewards/judge_quality/mean': 0.36249998211860657, 'rewards/judge_quality/std': 0.13562026619911194, 'rewards/repeat_penalty/mean': 0.43883436918258667, 'rewards/repeat_penalty/std': 0.17324909567832947, 'rewards/near_duplicate_penalty/mean': 0.8260517716407776, 'rewards/near_duplicate_penalty/std': 0.08417052030563354, 'rewards/opening_diversity/mean': 0.65625, 'rewards/opening_diversity/std': 0.1735912710428238, 'rewards/distinct_2/mean': 0.7871530055999756, 'rewards/distinct_2/std': 0.1044204831123352, 'rewards/total_composite/mean': 0.297556072473526, 'rewards/total_composite/std': 0.16931691765785217, 'reward': 0.297556072473526, 'reward_std': 0.16931691765785217, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0845629945397377, 'sampling/sampling_logp_difference/max': 1.9395453929901123, 'sampling/importance_sampling_ratio/min': 0.1437692940235138, 'sampling/importance_sampling_ratio/mean': 1.010769009590149, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3806200809776783, 'clip_ratio/low_mean': 0.009615384973585606, 'clip_ratio/low_min': 0.009615384973585606, 'clip_ratio/high_mean': 0.05670261709019542, 'clip_ratio/high_max': 0.05670261709019542, 'clip_ratio/region_mean': 0.06631800206378102, 'reward_total_mean': 0.297556072473526, 'reward_meter_mean': 0.7777613401412964, 'reward_meter_std': 0.34793820977211, '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_lexical_plausibility_mean': 0.9689165353775024, 'reward_lexical_plausibility_std': 0.0879172757267952, 'reward_repeat_penalty_mean': 0.43883436918258667, 'reward_repeat_penalty_std': 0.17324909567832947, 'reward_near_duplicate_penalty_mean': 0.8260517716407776, 'reward_near_duplicate_penalty_std': 0.08417052030563354, 'reward_opening_diversity_mean': 0.65625, 'reward_opening_diversity_std': 0.1735912710428238, 'reward_distinct_2_mean': 0.7871530055999756, 'reward_distinct_2_std': 0.1044204831123352, 'reward_judge_quality_mean': 0.36249998211860657, 'reward_judge_quality_std': 0.13562026619911194, 'reward_total_composite_mean': 0.297556072473526, 'reward_total_composite_std': 0.16931691765785217, 'epoch': 0.03}

 26%|██▌       | 866/3300 [3:29:05<8:23:02, 12.40s/it]INFO 04-12 14:15:17 [block_pool.py:378] Successfully reset prefix cache

 26%|██▋       | 867/3300 [3:29:18<8:38:44, 12.79s/it]2026-04-12 14:15:31,040 | INFO | train_grpo_train | metrics_logged mode=train step=867

                                                      
{'loss': -0.0947, 'grad_norm': 3.859180450439453, 'learning_rate': 7.375757575757576e-06, 'num_tokens': 1763915.0, 'completions/mean_length': 180.0, 'completions/min_length': 66.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 69.33333587646484, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.2914176285266876, 'rewards/meter/std': 0.15172293782234192, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9521939158439636, 'rewards/lexical_plausibility/std': 0.09111620485782623, 'rewards/judge_quality/mean': 0.48000001907348633, 'rewards/judge_quality/std': 0.31272992491722107, 'rewards/repeat_penalty/mean': 0.9737507104873657, 'rewards/repeat_penalty/std': 0.02878880500793457, 'rewards/near_duplicate_penalty/mean': 0.9803803563117981, 'rewards/near_duplicate_penalty/std': 0.015758827328681946, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9931318759918213, 'rewards/distinct_2/std': 0.01942601054906845, 'rewards/total_composite/mean': 0.16419778764247894, 'rewards/total_composite/std': 0.13353072106838226, 'reward': 0.16419778764247894, 'reward_std': 0.13353072106838226, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15073788166046143, 'sampling/sampling_logp_difference/max': 1.4994456768035889, 'sampling/importance_sampling_ratio/min': 0.22325389087200165, 'sampling/importance_sampling_ratio/mean': 1.0131865739822388, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5943051502108574, 'clip_ratio/low_mean': 0.03898390382528305, 'clip_ratio/low_min': 0.03898390382528305, 'clip_ratio/high_mean': 0.07402368821203709, 'clip_ratio/high_max': 0.07402368821203709, 'clip_ratio/region_mean': 0.11300759203732014, 'reward_total_mean': 0.16419778764247894, 'reward_meter_mean': 0.2914176285266876, 'reward_meter_std': 0.15172293782234192, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9521939158439636, 'reward_lexical_plausibility_std': 0.09111620485782623, 'reward_repeat_penalty_mean': 0.9737507104873657, 'reward_repeat_penalty_std': 0.02878880500793457, 'reward_near_duplicate_penalty_mean': 0.9803803563117981, 'reward_near_duplicate_penalty_std': 0.015758827328681946, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9931318759918213, 'reward_distinct_2_std': 0.01942601054906845, 'reward_judge_quality_mean': 0.48000001907348633, 'reward_judge_quality_std': 0.31272992491722107, 'reward_total_composite_mean': 0.16419778764247894, 'reward_total_composite_std': 0.13353072106838226, 'epoch': 0.03}

 26%|██▋       | 867/3300 [3:29:19<8:38:44, 12.79s/it]INFO 04-12 14:15:31 [block_pool.py:378] Successfully reset prefix cache

 26%|██▋       | 868/3300 [3:29:27<7:51:12, 11.63s/it]2026-04-12 14:15:39,935 | INFO | train_grpo_train | metrics_logged mode=train step=868

                                                      
{'loss': 0.05, 'grad_norm': 13.824076652526855, 'learning_rate': 7.372727272727274e-06, 'num_tokens': 1765763.0, 'completions/mean_length': 67.0, 'completions/min_length': 60.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.0, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.38709595799446106, 'rewards/meter/std': 0.3622923195362091, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.42499998211860657, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.9322495460510254, 'rewards/repeat_penalty/std': 0.1190701350569725, 'rewards/near_duplicate_penalty/mean': 0.9690883159637451, 'rewards/near_duplicate_penalty/std': 0.04162357375025749, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9587912559509277, 'rewards/distinct_2/std': 0.09065470844507217, 'rewards/total_composite/mean': 0.1587425172328949, 'rewards/total_composite/std': 0.14384205639362335, 'reward': 0.1587425172328949, 'reward_std': 0.14384204149246216, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1594058722257614, 'sampling/sampling_logp_difference/max': 2.085472583770752, 'sampling/importance_sampling_ratio/min': 0.12424838542938232, 'sampling/importance_sampling_ratio/mean': 0.9888842701911926, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6814248487353325, 'clip_ratio/low_mean': 0.07203493919223547, 'clip_ratio/low_min': 0.07203493919223547, 'clip_ratio/high_mean': 0.04719291161745787, 'clip_ratio/high_max': 0.04719291161745787, 'clip_ratio/region_mean': 0.11922785080969334, 'reward_total_mean': 0.1587425172328949, 'reward_meter_mean': 0.38709595799446106, 'reward_meter_std': 0.3622923195362091, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.9322495460510254, 'reward_repeat_penalty_std': 0.1190701350569725, 'reward_near_duplicate_penalty_mean': 0.9690883159637451, 'reward_near_duplicate_penalty_std': 0.04162357375025749, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9587912559509277, 'reward_distinct_2_std': 0.09065470844507217, 'reward_judge_quality_mean': 0.42499998211860657, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.1587425172328949, 'reward_total_composite_std': 0.14384205639362335, 'epoch': 0.03}

 26%|██▋       | 868/3300 [3:29:27<7:51:12, 11.63s/it]INFO 04-12 14:15:40 [block_pool.py:378] Successfully reset prefix cache

 26%|██▋       | 869/3300 [3:29:41<8:11:13, 12.12s/it]2026-04-12 14:15:53,226 | INFO | train_grpo_train | metrics_logged mode=train step=869

                                                      
{'loss': -0.0813, 'grad_norm': 3.225282669067383, 'learning_rate': 7.36969696969697e-06, 'num_tokens': 1767384.0, 'completions/mean_length': 177.625, 'completions/min_length': 46.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 66.16667175292969, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.36734604835510254, 'rewards/meter/std': 0.3179638683795929, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9347068071365356, 'rewards/lexical_plausibility/std': 0.12127043306827545, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.21380898356437683, 'rewards/repeat_penalty/mean': 0.6754549145698547, 'rewards/repeat_penalty/std': 0.21429799497127533, 'rewards/near_duplicate_penalty/mean': 0.8525331020355225, 'rewards/near_duplicate_penalty/std': 0.08067642897367477, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.8559688329696655, 'rewards/distinct_2/std': 0.12709783017635345, 'rewards/total_composite/mean': 0.1239505410194397, 'rewards/total_composite/std': 0.10338914394378662, 'reward': 0.1239505410194397, 'reward_std': 0.10338913649320602, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13509614765644073, 'sampling/sampling_logp_difference/max': 1.2986257076263428, 'sampling/importance_sampling_ratio/min': 0.2854008674621582, 'sampling/importance_sampling_ratio/mean': 1.001600742340088, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.45891711115837097, 'clip_ratio/low_mean': 0.026167278178036213, 'clip_ratio/low_min': 0.026167278178036213, 'clip_ratio/high_mean': 0.05937543511390686, 'clip_ratio/high_max': 0.05937543511390686, 'clip_ratio/region_mean': 0.08554271329194307, 'reward_total_mean': 0.1239505410194397, 'reward_meter_mean': 0.36734604835510254, 'reward_meter_std': 0.3179638683795929, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9347068071365356, 'reward_lexical_plausibility_std': 0.12127043306827545, 'reward_repeat_penalty_mean': 0.6754549145698547, 'reward_repeat_penalty_std': 0.21429799497127533, 'reward_near_duplicate_penalty_mean': 0.8525331020355225, 'reward_near_duplicate_penalty_std': 0.08067642897367477, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.8559688329696655, 'reward_distinct_2_std': 0.12709783017635345, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.21380898356437683, 'reward_total_composite_mean': 0.1239505410194397, 'reward_total_composite_std': 0.10338914394378662, 'epoch': 0.03}

 26%|██▋       | 869/3300 [3:29:41<8:11:13, 12.12s/it]INFO 04-12 14:15:53 [block_pool.py:378] Successfully reset prefix cache

 26%|██▋       | 870/3300 [3:29:50<7:42:06, 11.41s/it]2026-04-12 14:16:02,977 | INFO | train_grpo_train | metrics_logged mode=train step=870

                                                      
{'loss': 0.0215, 'grad_norm': 7.330625534057617, 'learning_rate': 7.3666666666666676e-06, 'num_tokens': 1769847.0, 'completions/mean_length': 126.875, 'completions/min_length': 112.0, 'completions/max_length': 149.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 126.875, 'completions/min_terminated_length': 112.0, 'completions/max_terminated_length': 149.0, 'rewards/meter/mean': 0.8272303342819214, 'rewards/meter/std': 0.18839241564273834, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3375000059604645, 'rewards/judge_quality/std': 0.08345229178667068, 'rewards/repeat_penalty/mean': 0.5826253294944763, 'rewards/repeat_penalty/std': 0.16902408003807068, 'rewards/near_duplicate_penalty/mean': 0.8882546424865723, 'rewards/near_duplicate_penalty/std': 0.03844599425792694, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.708530843257904, 'rewards/distinct_2/std': 0.15559612214565277, 'rewards/total_composite/mean': 0.2730218172073364, 'rewards/total_composite/std': 0.10799101740121841, 'reward': 0.2730218172073364, 'reward_std': 0.10799100995063782, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10582651197910309, 'sampling/sampling_logp_difference/max': 2.2159557342529297, 'sampling/importance_sampling_ratio/min': 0.1090492457151413, 'sampling/importance_sampling_ratio/mean': 1.0091804265975952, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4706226587295532, 'clip_ratio/low_mean': 0.05683693056926131, 'clip_ratio/low_min': 0.05683693056926131, 'clip_ratio/high_mean': 0.04243682138621807, 'clip_ratio/high_max': 0.04243682138621807, 'clip_ratio/region_mean': 0.09927375195547938, 'reward_total_mean': 0.2730218172073364, 'reward_meter_mean': 0.8272303342819214, 'reward_meter_std': 0.18839241564273834, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.09258200973272324, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.5826253294944763, 'reward_repeat_penalty_std': 0.16902408003807068, 'reward_near_duplicate_penalty_mean': 0.8882546424865723, 'reward_near_duplicate_penalty_std': 0.03844599425792694, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.708530843257904, 'reward_distinct_2_std': 0.15559612214565277, 'reward_judge_quality_mean': 0.3375000059604645, 'reward_judge_quality_std': 0.08345229178667068, 'reward_total_composite_mean': 0.2730218172073364, 'reward_total_composite_std': 0.10799101740121841, 'epoch': 0.03}

 26%|██▋       | 870/3300 [3:29:50<7:42:06, 11.41s/it]INFO 04-12 14:16:03 [block_pool.py:378] Successfully reset prefix cache

 26%|██▋       | 871/3300 [3:30:04<8:07:46, 12.05s/it]2026-04-12 14:16:16,511 | INFO | train_grpo_train | metrics_logged mode=train step=871

                                                      
{'loss': -0.0619, 'grad_norm': 6.001213550567627, 'learning_rate': 7.363636363636364e-06, 'num_tokens': 1772137.0, 'completions/mean_length': 175.25, 'completions/min_length': 112.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 127.14286041259766, 'completions/min_terminated_length': 112.0, 'completions/max_terminated_length': 157.0, 'rewards/meter/mean': 0.39989596605300903, 'rewards/meter/std': 0.3121834397315979, 'rewards/count_adherence/mean': 0.8958333134651184, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9654213190078735, 'rewards/lexical_plausibility/std': 0.09780332446098328, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.7331823110580444, 'rewards/repeat_penalty/std': 0.22717273235321045, 'rewards/near_duplicate_penalty/mean': 0.9222255945205688, 'rewards/near_duplicate_penalty/std': 0.05219656229019165, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.7962535619735718, 'rewards/distinct_2/std': 0.2072196751832962, 'rewards/total_composite/mean': 0.08924335241317749, 'rewards/total_composite/std': 0.10416724532842636, 'reward': 0.08924335241317749, 'reward_std': 0.10416723787784576, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1451282948255539, 'sampling/sampling_logp_difference/max': 2.7776942253112793, 'sampling/importance_sampling_ratio/min': 0.06218171864748001, 'sampling/importance_sampling_ratio/mean': 0.9980655312538147, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.578022550791502, 'clip_ratio/low_mean': 0.08239227440208197, 'clip_ratio/low_min': 0.08239227440208197, 'clip_ratio/high_mean': 0.0257191713899374, 'clip_ratio/high_max': 0.0257191713899374, 'clip_ratio/region_mean': 0.10811144579201937, 'reward_total_mean': 0.08924335241317749, 'reward_meter_mean': 0.39989596605300903, 'reward_meter_std': 0.3121834397315979, 'reward_count_adherence_mean': 0.8958333134651184, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9654213190078735, 'reward_lexical_plausibility_std': 0.09780332446098328, 'reward_repeat_penalty_mean': 0.7331823110580444, 'reward_repeat_penalty_std': 0.22717273235321045, 'reward_near_duplicate_penalty_mean': 0.9222255945205688, 'reward_near_duplicate_penalty_std': 0.05219656229019165, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.7962535619735718, 'reward_distinct_2_std': 0.2072196751832962, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.08924335241317749, 'reward_total_composite_std': 0.10416724532842636, 'epoch': 0.03}

 26%|██▋       | 871/3300 [3:30:04<8:07:46, 12.05s/it]INFO 04-12 14:16:16 [block_pool.py:378] Successfully reset prefix cache

 26%|██▋       | 872/3300 [3:30:14<7:49:17, 11.60s/it]2026-04-12 14:16:27,050 | INFO | train_grpo_train | metrics_logged mode=train step=872

                                                      
{'loss': -0.0197, 'grad_norm': 7.632763862609863, 'learning_rate': 7.360606060606061e-06, 'num_tokens': 1774674.0, 'completions/mean_length': 127.125, 'completions/min_length': 105.0, 'completions/max_length': 141.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 127.125, 'completions/min_terminated_length': 105.0, 'completions/max_terminated_length': 141.0, 'rewards/meter/mean': 0.5476891994476318, 'rewards/meter/std': 0.34833618998527527, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.8275880813598633, 'rewards/repeat_penalty/std': 0.10005103796720505, 'rewards/near_duplicate_penalty/mean': 0.9418987035751343, 'rewards/near_duplicate_penalty/std': 0.013158251531422138, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8921672105789185, 'rewards/distinct_2/std': 0.07349219173192978, 'rewards/total_composite/mean': 0.20300811529159546, 'rewards/total_composite/std': 0.13897737860679626, 'reward': 0.20300811529159546, 'reward_std': 0.13897736370563507, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11417630314826965, 'sampling/sampling_logp_difference/max': 2.6972084045410156, 'sampling/importance_sampling_ratio/min': 0.06739338487386703, 'sampling/importance_sampling_ratio/mean': 0.9951967000961304, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5607620626688004, 'clip_ratio/low_mean': 0.05955591704696417, 'clip_ratio/low_min': 0.05955591704696417, 'clip_ratio/high_mean': 0.05671980790793896, 'clip_ratio/high_max': 0.05671980790793896, 'clip_ratio/region_mean': 0.11627572495490313, 'reward_total_mean': 0.20300811529159546, 'reward_meter_mean': 0.5476891994476318, 'reward_meter_std': 0.34833618998527527, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8275880813598633, 'reward_repeat_penalty_std': 0.10005103796720505, 'reward_near_duplicate_penalty_mean': 0.9418987035751343, 'reward_near_duplicate_penalty_std': 0.013158251531422138, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8921672105789185, 'reward_distinct_2_std': 0.07349219173192978, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.20300811529159546, 'reward_total_composite_std': 0.13897737860679626, 'epoch': 0.04}

 26%|██▋       | 872/3300 [3:30:15<7:49:17, 11.60s/it]INFO 04-12 14:16:27 [block_pool.py:378] Successfully reset prefix cache

 26%|██▋       | 873/3300 [3:30:28<8:14:57, 12.24s/it]2026-04-12 14:16:40,786 | INFO | train_grpo_train | metrics_logged mode=train step=873

                                                      
{'loss': -0.0555, 'grad_norm': 4.501315593719482, 'learning_rate': 7.357575757575758e-06, 'num_tokens': 1776302.0, 'completions/mean_length': 112.5, 'completions/min_length': 51.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 55.42857360839844, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.8306863307952881, 'rewards/meter/std': 0.33223599195480347, '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/lexical_plausibility/mean': 0.9848498106002808, 'rewards/lexical_plausibility/std': 0.04285118728876114, 'rewards/judge_quality/mean': 0.3462499976158142, 'rewards/judge_quality/std': 0.2840491831302643, 'rewards/repeat_penalty/mean': 0.8906967639923096, 'rewards/repeat_penalty/std': 0.12793321907520294, 'rewards/near_duplicate_penalty/mean': 0.9242289066314697, 'rewards/near_duplicate_penalty/std': 0.08066439628601074, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9602006673812866, 'rewards/distinct_2/std': 0.07493601739406586, 'rewards/total_composite/mean': 0.3201238512992859, 'rewards/total_composite/std': 0.2824220061302185, 'reward': 0.3201238512992859, 'reward_std': 0.2824220061302185, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13651838898658752, 'sampling/sampling_logp_difference/max': 2.6320786476135254, 'sampling/importance_sampling_ratio/min': 0.07192879170179367, 'sampling/importance_sampling_ratio/mean': 1.008743405342102, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6360415667295456, 'clip_ratio/low_mean': 0.035882867872714996, 'clip_ratio/low_min': 0.035882867872714996, 'clip_ratio/high_mean': 0.06075863540172577, 'clip_ratio/high_max': 0.06075863540172577, 'clip_ratio/region_mean': 0.09664150327444077, 'reward_total_mean': 0.3201238512992859, 'reward_meter_mean': 0.8306863307952881, 'reward_meter_std': 0.33223599195480347, '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_lexical_plausibility_mean': 0.9848498106002808, 'reward_lexical_plausibility_std': 0.04285118728876114, 'reward_repeat_penalty_mean': 0.8906967639923096, 'reward_repeat_penalty_std': 0.12793321907520294, 'reward_near_duplicate_penalty_mean': 0.9242289066314697, 'reward_near_duplicate_penalty_std': 0.08066439628601074, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9602006673812866, 'reward_distinct_2_std': 0.07493601739406586, 'reward_judge_quality_mean': 0.3462499976158142, 'reward_judge_quality_std': 0.2840491831302643, 'reward_total_composite_mean': 0.3201238512992859, 'reward_total_composite_std': 0.2824220061302185, 'epoch': 0.04}

 26%|██▋       | 873/3300 [3:30:28<8:14:57, 12.24s/it]INFO 04-12 14:16:41 [block_pool.py:378] Successfully reset prefix cache

 26%|██▋       | 874/3300 [3:30:37<7:37:54, 11.33s/it]2026-04-12 14:16:49,982 | INFO | train_grpo_train | metrics_logged mode=train step=874

                                                      
{'loss': 0.1414, 'grad_norm': 7.832292079925537, 'learning_rate': 7.354545454545456e-06, 'num_tokens': 1778317.0, 'completions/mean_length': 79.875, 'completions/min_length': 56.0, 'completions/max_length': 105.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.875, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.29746511578559875, 'rewards/meter/std': 0.2942606508731842, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.09258200973272324, 'rewards/repeat_penalty/mean': 0.5523877143859863, 'rewards/repeat_penalty/std': 0.2730301022529602, 'rewards/near_duplicate_penalty/mean': 0.8659970164299011, 'rewards/near_duplicate_penalty/std': 0.1328260600566864, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.1602174937725067, 'rewards/distinct_2/mean': 0.7535686492919922, 'rewards/distinct_2/std': 0.20347315073013306, 'rewards/total_composite/mean': 0.09182830154895782, 'rewards/total_composite/std': 0.06809425354003906, 'reward': 0.09182830154895782, 'reward_std': 0.06809425354003906, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10342110693454742, 'sampling/sampling_logp_difference/max': 3.3455324172973633, 'sampling/importance_sampling_ratio/min': 0.035241447389125824, 'sampling/importance_sampling_ratio/mean': 1.009207844734192, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5651168003678322, 'clip_ratio/low_mean': 0.05296321352943778, 'clip_ratio/low_min': 0.05296321352943778, 'clip_ratio/high_mean': 0.03530844207853079, 'clip_ratio/high_max': 0.03530844207853079, 'clip_ratio/region_mean': 0.08827165560796857, 'reward_total_mean': 0.09182830154895782, 'reward_meter_mean': 0.29746511578559875, 'reward_meter_std': 0.2942606508731842, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.5523877143859863, 'reward_repeat_penalty_std': 0.2730301022529602, 'reward_near_duplicate_penalty_mean': 0.8659970164299011, 'reward_near_duplicate_penalty_std': 0.1328260600566864, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.1602174937725067, 'reward_distinct_2_mean': 0.7535686492919922, 'reward_distinct_2_std': 0.20347315073013306, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.09258200973272324, 'reward_total_composite_mean': 0.09182830154895782, 'reward_total_composite_std': 0.06809425354003906, 'epoch': 0.04}

 26%|██▋       | 874/3300 [3:30:37<7:37:54, 11.33s/it]INFO 04-12 14:16:50 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 875/3300 [3:30:46<7:08:00, 10.59s/it]2026-04-12 14:16:58,862 | INFO | train_grpo_train | metrics_logged mode=train step=875

                                                      
{'loss': 0.0711, 'grad_norm': 10.09304428100586, 'learning_rate': 7.351515151515151e-06, 'num_tokens': 1780299.0, 'completions/mean_length': 71.75, 'completions/min_length': 63.0, 'completions/max_length': 83.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.75, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.9014970660209656, 'rewards/meter/std': 0.13068506121635437, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4375, 'rewards/judge_quality/std': 0.13562028110027313, 'rewards/repeat_penalty/mean': 0.8311077356338501, 'rewards/repeat_penalty/std': 0.11203678697347641, 'rewards/near_duplicate_penalty/mean': 0.9320562481880188, 'rewards/near_duplicate_penalty/std': 0.028182754293084145, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8894095420837402, 'rewards/distinct_2/std': 0.09910806268453598, 'rewards/total_composite/mean': 0.38931509852409363, 'rewards/total_composite/std': 0.11298346519470215, 'reward': 0.38931509852409363, 'reward_std': 0.11298347264528275, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1445842832326889, 'sampling/sampling_logp_difference/max': 1.7441703081130981, 'sampling/importance_sampling_ratio/min': 0.17478995025157928, 'sampling/importance_sampling_ratio/mean': 1.0147509574890137, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.779978558421135, 'clip_ratio/low_mean': 0.08296741731464863, 'clip_ratio/low_min': 0.08296741731464863, 'clip_ratio/high_mean': 0.060111574828624725, 'clip_ratio/high_max': 0.060111574828624725, 'clip_ratio/region_mean': 0.14307899214327335, 'reward_total_mean': 0.38931509852409363, 'reward_meter_mean': 0.9014970660209656, 'reward_meter_std': 0.13068506121635437, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8311077356338501, 'reward_repeat_penalty_std': 0.11203678697347641, 'reward_near_duplicate_penalty_mean': 0.9320562481880188, 'reward_near_duplicate_penalty_std': 0.028182754293084145, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8894095420837402, 'reward_distinct_2_std': 0.09910806268453598, 'reward_judge_quality_mean': 0.4375, 'reward_judge_quality_std': 0.13562028110027313, 'reward_total_composite_mean': 0.38931509852409363, 'reward_total_composite_std': 0.11298346519470215, 'epoch': 0.04}

 27%|██▋       | 875/3300 [3:30:46<7:08:00, 10.59s/it]INFO 04-12 14:16:59 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 876/3300 [3:30:58<7:21:50, 10.94s/it]2026-04-12 14:17:10,602 | INFO | train_grpo_train | metrics_logged mode=train step=876

                                                      
{'loss': 0.0299, 'grad_norm': 4.762448310852051, 'learning_rate': 7.348484848484849e-06, 'num_tokens': 1783902.0, 'completions/mean_length': 236.375, 'completions/min_length': 208.0, 'completions/max_length': 280.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 236.375, 'completions/min_terminated_length': 208.0, 'completions/max_terminated_length': 280.0, 'rewards/meter/mean': 0.5270986557006836, 'rewards/meter/std': 0.2685558795928955, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.06681530922651291, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.28528380393981934, 'rewards/repeat_penalty/std': 0.19556787610054016, 'rewards/near_duplicate_penalty/mean': 0.8050830364227295, 'rewards/near_duplicate_penalty/std': 0.06882479786872864, 'rewards/opening_diversity/mean': 0.9963235259056091, 'rewards/opening_diversity/std': 0.010398639366030693, 'rewards/distinct_2/mean': 0.38537976145744324, 'rewards/distinct_2/std': 0.2111949622631073, 'rewards/total_composite/mean': 0.13332274556159973, 'rewards/total_composite/std': 0.07764926552772522, 'reward': 0.13332274556159973, 'reward_std': 0.07764925807714462, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09145982563495636, 'sampling/sampling_logp_difference/max': 2.0879368782043457, 'sampling/importance_sampling_ratio/min': 0.12394257634878159, 'sampling/importance_sampling_ratio/mean': 1.0108877420425415, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4582247845828533, 'clip_ratio/low_mean': 0.0729204248636961, 'clip_ratio/low_min': 0.0729204248636961, 'clip_ratio/high_mean': 0.0327193234115839, 'clip_ratio/high_max': 0.0327193234115839, 'clip_ratio/region_mean': 0.10563974827528, 'reward_total_mean': 0.13332274556159973, 'reward_meter_mean': 0.5270986557006836, 'reward_meter_std': 0.2685558795928955, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.06681530922651291, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.28528380393981934, 'reward_repeat_penalty_std': 0.19556787610054016, 'reward_near_duplicate_penalty_mean': 0.8050830364227295, 'reward_near_duplicate_penalty_std': 0.06882479786872864, 'reward_opening_diversity_mean': 0.9963235259056091, 'reward_opening_diversity_std': 0.010398639366030693, 'reward_distinct_2_mean': 0.38537976145744324, 'reward_distinct_2_std': 0.2111949622631073, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.13332274556159973, 'reward_total_composite_std': 0.07764926552772522, 'epoch': 0.04}

 27%|██▋       | 876/3300 [3:30:58<7:21:50, 10.94s/it]INFO 04-12 14:17:10 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 877/3300 [3:31:06<6:45:11, 10.03s/it]2026-04-12 14:17:18,527 | INFO | train_grpo_train | metrics_logged mode=train step=877

                                                      
{'loss': 0.0469, 'grad_norm': 19.530975341796875, 'learning_rate': 7.345454545454546e-06, 'num_tokens': 1785613.0, 'completions/mean_length': 56.875, 'completions/min_length': 55.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.875, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.8635131120681763, 'rewards/meter/std': 0.2633155584335327, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.48875001072883606, 'rewards/judge_quality/std': 0.3626662492752075, 'rewards/repeat_penalty/mean': 0.8042378425598145, 'rewards/repeat_penalty/std': 0.09148935973644257, 'rewards/near_duplicate_penalty/mean': 0.9202189445495605, 'rewards/near_duplicate_penalty/std': 0.05062742531299591, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8718816041946411, 'rewards/distinct_2/std': 0.06363079696893692, 'rewards/total_composite/mean': 0.4415525197982788, 'rewards/total_composite/std': 0.34574729204177856, 'reward': 0.4415525197982788, 'reward_std': 0.34574729204177856, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09606894105672836, 'sampling/sampling_logp_difference/max': 1.6975698471069336, 'sampling/importance_sampling_ratio/min': 0.1831280142068863, 'sampling/importance_sampling_ratio/mean': 0.9880635142326355, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4432758465409279, 'clip_ratio/low_mean': 0.06798952259123325, 'clip_ratio/low_min': 0.06798952259123325, 'clip_ratio/high_mean': 0.033223243430256844, 'clip_ratio/high_max': 0.033223243430256844, 'clip_ratio/region_mean': 0.1012127660214901, 'reward_total_mean': 0.4415525197982788, 'reward_meter_mean': 0.8635131120681763, 'reward_meter_std': 0.2633155584335327, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8042378425598145, 'reward_repeat_penalty_std': 0.09148935973644257, 'reward_near_duplicate_penalty_mean': 0.9202189445495605, 'reward_near_duplicate_penalty_std': 0.05062742531299591, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8718816041946411, 'reward_distinct_2_std': 0.06363079696893692, 'reward_judge_quality_mean': 0.48875001072883606, 'reward_judge_quality_std': 0.3626662492752075, 'reward_total_composite_mean': 0.4415525197982788, 'reward_total_composite_std': 0.34574729204177856, 'epoch': 0.04}

 27%|██▋       | 877/3300 [3:31:06<6:45:11, 10.03s/it]INFO 04-12 14:17:18 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 878/3300 [3:31:20<7:31:45, 11.19s/it]2026-04-12 14:17:32,413 | INFO | train_grpo_train | metrics_logged mode=train step=878

                                                      
{'loss': -0.0678, 'grad_norm': 2.321756601333618, 'learning_rate': 7.342424242424243e-06, 'num_tokens': 1787140.0, 'completions/mean_length': 223.875, 'completions/min_length': 48.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 51.0, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.659151554107666, 'rewards/meter/std': 0.4500230550765991, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.8951897025108337, 'rewards/lexical_plausibility/std': 0.1461409479379654, 'rewards/judge_quality/mean': 0.29625001549720764, 'rewards/judge_quality/std': 0.2980382442474365, 'rewards/repeat_penalty/mean': 0.7111141681671143, 'rewards/repeat_penalty/std': 0.29982367157936096, 'rewards/near_duplicate_penalty/mean': 0.8549773693084717, 'rewards/near_duplicate_penalty/std': 0.16954217851161957, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.8347539305686951, 'rewards/distinct_2/std': 0.23982471227645874, 'rewards/total_composite/mean': 0.27332350611686707, 'rewards/total_composite/std': 0.31190598011016846, 'reward': 0.27332350611686707, 'reward_std': 0.31190595030784607, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11236979067325592, 'sampling/sampling_logp_difference/max': 1.2867416143417358, 'sampling/importance_sampling_ratio/min': 0.27616918087005615, 'sampling/importance_sampling_ratio/mean': 0.9936332702636719, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.37248532474040985, 'clip_ratio/low_mean': 0.010775862261652946, 'clip_ratio/low_min': 0.010775862261652946, 'clip_ratio/high_mean': 0.07108418364077806, 'clip_ratio/high_max': 0.07108418364077806, 'clip_ratio/region_mean': 0.08186004590243101, 'reward_total_mean': 0.27332350611686707, 'reward_meter_mean': 0.659151554107666, 'reward_meter_std': 0.4500230550765991, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.8951897025108337, 'reward_lexical_plausibility_std': 0.1461409479379654, 'reward_repeat_penalty_mean': 0.7111141681671143, 'reward_repeat_penalty_std': 0.29982367157936096, 'reward_near_duplicate_penalty_mean': 0.8549773693084717, 'reward_near_duplicate_penalty_std': 0.16954217851161957, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.8347539305686951, 'reward_distinct_2_std': 0.23982471227645874, 'reward_judge_quality_mean': 0.29625001549720764, 'reward_judge_quality_std': 0.2980382442474365, 'reward_total_composite_mean': 0.27332350611686707, 'reward_total_composite_std': 0.31190598011016846, 'epoch': 0.04}

 27%|██▋       | 878/3300 [3:31:20<7:31:45, 11.19s/it]INFO 04-12 14:17:32 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 879/3300 [3:31:34<8:03:40, 11.99s/it]2026-04-12 14:17:46,259 | INFO | train_grpo_train | metrics_logged mode=train step=879

                                                      
{'loss': -0.0941, 'grad_norm': 5.5998148918151855, 'learning_rate': 7.3393939393939395e-06, 'num_tokens': 1788731.0, 'completions/mean_length': 106.875, 'completions/min_length': 42.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 49.000003814697266, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.7215713858604431, 'rewards/meter/std': 0.4185429513454437, '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/lexical_plausibility/mean': 0.9464878439903259, 'rewards/lexical_plausibility/std': 0.10046496242284775, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.14078859984874725, 'rewards/repeat_penalty/mean': 0.8929903507232666, 'rewards/repeat_penalty/std': 0.08757708221673965, 'rewards/near_duplicate_penalty/mean': 0.912240743637085, 'rewards/near_duplicate_penalty/std': 0.06082136183977127, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9773755669593811, 'rewards/distinct_2/std': 0.041892312467098236, 'rewards/total_composite/mean': 0.30992770195007324, 'rewards/total_composite/std': 0.18913213908672333, 'reward': 0.30992770195007324, 'reward_std': 0.18913212418556213, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16873911023139954, 'sampling/sampling_logp_difference/max': 3.967195749282837, 'sampling/importance_sampling_ratio/min': 0.018926434218883514, 'sampling/importance_sampling_ratio/mean': 0.9758078455924988, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5684429481625557, 'clip_ratio/low_mean': 0.005319148767739534, 'clip_ratio/low_min': 0.005319148767739534, 'clip_ratio/high_mean': 0.1118007549084723, 'clip_ratio/high_max': 0.1118007549084723, 'clip_ratio/region_mean': 0.11711990367621183, 'reward_total_mean': 0.30992770195007324, 'reward_meter_mean': 0.7215713858604431, 'reward_meter_std': 0.4185429513454437, '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_lexical_plausibility_mean': 0.9464878439903259, 'reward_lexical_plausibility_std': 0.10046496242284775, 'reward_repeat_penalty_mean': 0.8929903507232666, 'reward_repeat_penalty_std': 0.08757708221673965, 'reward_near_duplicate_penalty_mean': 0.912240743637085, 'reward_near_duplicate_penalty_std': 0.06082136183977127, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9773755669593811, 'reward_distinct_2_std': 0.041892312467098236, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.14078859984874725, 'reward_total_composite_mean': 0.30992770195007324, 'reward_total_composite_std': 0.18913213908672333, 'epoch': 0.04}

 27%|██▋       | 879/3300 [3:31:34<8:03:40, 11.99s/it]INFO 04-12 14:17:46 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 880/3300 [3:31:42<7:19:41, 10.90s/it]2026-04-12 14:17:54,630 | INFO | train_grpo_train | metrics_logged mode=train step=880

                                                      
{'loss': 0.0097, 'grad_norm': 10.49024772644043, 'learning_rate': 7.336363636363637e-06, 'num_tokens': 1790709.0, 'completions/mean_length': 74.25, 'completions/min_length': 65.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.25, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.5394495725631714, 'rewards/meter/std': 0.3519377112388611, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9955357313156128, 'rewards/lexical_plausibility/std': 0.008266246877610683, 'rewards/judge_quality/mean': 0.23750001192092896, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.7814996242523193, 'rewards/repeat_penalty/std': 0.11847735941410065, 'rewards/near_duplicate_penalty/mean': 0.9319391250610352, 'rewards/near_duplicate_penalty/std': 0.05879274010658264, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8353081345558167, 'rewards/distinct_2/std': 0.09324527531862259, 'rewards/total_composite/mean': 0.12256370484828949, 'rewards/total_composite/std': 0.11397451162338257, 'reward': 0.12256370484828949, 'reward_std': 0.11397451162338257, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.132102370262146, 'sampling/sampling_logp_difference/max': 2.2229855060577393, 'sampling/importance_sampling_ratio/min': 0.1082853376865387, 'sampling/importance_sampling_ratio/mean': 1.0134063959121704, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7012450322508812, 'clip_ratio/low_mean': 0.09132364299148321, 'clip_ratio/low_min': 0.09132364299148321, 'clip_ratio/high_mean': 0.03963879402726889, 'clip_ratio/high_max': 0.03963879402726889, 'clip_ratio/region_mean': 0.1309624370187521, 'reward_total_mean': 0.12256370484828949, 'reward_meter_mean': 0.5394495725631714, 'reward_meter_std': 0.3519377112388611, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9955357313156128, 'reward_lexical_plausibility_std': 0.008266246877610683, 'reward_repeat_penalty_mean': 0.7814996242523193, 'reward_repeat_penalty_std': 0.11847735941410065, 'reward_near_duplicate_penalty_mean': 0.9319391250610352, 'reward_near_duplicate_penalty_std': 0.05879274010658264, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8353081345558167, 'reward_distinct_2_std': 0.09324527531862259, 'reward_judge_quality_mean': 0.23750001192092896, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.12256370484828949, 'reward_total_composite_std': 0.11397451162338257, 'epoch': 0.04}

 27%|██▋       | 880/3300 [3:31:42<7:19:41, 10.90s/it]INFO 04-12 14:17:54 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 881/3300 [3:31:56<7:58:51, 11.88s/it]2026-04-12 14:18:08,802 | INFO | train_grpo_train | metrics_logged mode=train step=881

                                                      
{'loss': -0.0076, 'grad_norm': 2.9933884143829346, 'learning_rate': 7.333333333333333e-06, 'num_tokens': 1795458.0, 'completions/mean_length': 440.625, 'completions/min_length': 372.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 430.4285888671875, 'completions/min_terminated_length': 372.0, 'completions/max_terminated_length': 486.0, 'rewards/meter/mean': 0.23844513297080994, 'rewards/meter/std': 0.2061980962753296, 'rewards/count_adherence/mean': 0.7727273106575012, 'rewards/count_adherence/std': 0.06872081011533737, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.037745535373687744, 'rewards/repeat_penalty/std': 0.05902951583266258, 'rewards/near_duplicate_penalty/mean': 0.7385398149490356, 'rewards/near_duplicate_penalty/std': 0.11261305212974548, 'rewards/opening_diversity/mean': 0.9590517282485962, 'rewards/opening_diversity/std': 0.11581922322511673, 'rewards/distinct_2/mean': 0.0638369545340538, 'rewards/distinct_2/std': 0.10444370657205582, 'rewards/total_composite/mean': 0.02498304285109043, 'rewards/total_composite/std': 0.02411171793937683, 'reward': 0.02498304285109043, 'reward_std': 0.024111716076731682, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05810955911874771, 'sampling/sampling_logp_difference/max': 2.770911693572998, 'sampling/importance_sampling_ratio/min': 0.06260490417480469, 'sampling/importance_sampling_ratio/mean': 0.9986941814422607, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2401742935180664, 'clip_ratio/low_mean': 0.02730027073994279, 'clip_ratio/low_min': 0.02730027073994279, 'clip_ratio/high_mean': 0.018933712039142847, 'clip_ratio/high_max': 0.018933712039142847, 'clip_ratio/region_mean': 0.046233982779085636, 'reward_total_mean': 0.02498304285109043, 'reward_meter_mean': 0.23844513297080994, 'reward_meter_std': 0.2061980962753296, 'reward_count_adherence_mean': 0.7727273106575012, 'reward_count_adherence_std': 0.06872081011533737, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.037745535373687744, 'reward_repeat_penalty_std': 0.05902951583266258, 'reward_near_duplicate_penalty_mean': 0.7385398149490356, 'reward_near_duplicate_penalty_std': 0.11261305212974548, 'reward_opening_diversity_mean': 0.9590517282485962, 'reward_opening_diversity_std': 0.11581922322511673, 'reward_distinct_2_mean': 0.0638369545340538, 'reward_distinct_2_std': 0.10444370657205582, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.02498304285109043, 'reward_total_composite_std': 0.02411171793937683, 'epoch': 0.04}

 27%|██▋       | 881/3300 [3:31:56<7:58:51, 11.88s/it]INFO 04-12 14:18:09 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 882/3300 [3:32:10<8:22:26, 12.47s/it]2026-04-12 14:18:22,630 | INFO | train_grpo_train | metrics_logged mode=train step=882

                                                      
{'loss': -0.1186, 'grad_norm': 4.406069755554199, 'learning_rate': 7.330303030303031e-06, 'num_tokens': 1797942.0, 'completions/mean_length': 173.5, 'completions/min_length': 118.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 125.14286041259766, 'completions/min_terminated_length': 118.0, 'completions/max_terminated_length': 134.0, 'rewards/meter/mean': 0.6663212180137634, 'rewards/meter/std': 0.3866715729236603, '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/lexical_plausibility/mean': 0.9844070672988892, 'rewards/lexical_plausibility/std': 0.044103577733039856, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.13562026619911194, 'rewards/repeat_penalty/mean': 0.6938369870185852, 'rewards/repeat_penalty/std': 0.1847635954618454, 'rewards/near_duplicate_penalty/mean': 0.8913737535476685, 'rewards/near_duplicate_penalty/std': 0.06611061096191406, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.7686411738395691, 'rewards/distinct_2/std': 0.16424210369586945, 'rewards/total_composite/mean': 0.27357497811317444, 'rewards/total_composite/std': 0.18653017282485962, 'reward': 0.27357497811317444, 'reward_std': 0.18653017282485962, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10844583809375763, 'sampling/sampling_logp_difference/max': 2.675227642059326, 'sampling/importance_sampling_ratio/min': 0.06889114528894424, 'sampling/importance_sampling_ratio/mean': 1.0037729740142822, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.373327162116766, 'clip_ratio/low_mean': 0.019054178148508072, 'clip_ratio/low_min': 0.019054178148508072, 'clip_ratio/high_mean': 0.06575033254921436, 'clip_ratio/high_max': 0.06575033254921436, 'clip_ratio/region_mean': 0.08480451069772243, 'reward_total_mean': 0.27357497811317444, 'reward_meter_mean': 0.6663212180137634, 'reward_meter_std': 0.3866715729236603, '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_lexical_plausibility_mean': 0.9844070672988892, 'reward_lexical_plausibility_std': 0.044103577733039856, 'reward_repeat_penalty_mean': 0.6938369870185852, 'reward_repeat_penalty_std': 0.1847635954618454, 'reward_near_duplicate_penalty_mean': 0.8913737535476685, 'reward_near_duplicate_penalty_std': 0.06611061096191406, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.7686411738395691, 'reward_distinct_2_std': 0.16424210369586945, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.13562026619911194, 'reward_total_composite_mean': 0.27357497811317444, 'reward_total_composite_std': 0.18653017282485962, 'epoch': 0.04}

 27%|██▋       | 882/3300 [3:32:10<8:22:26, 12.47s/it]INFO 04-12 14:18:22 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 883/3300 [3:32:24<8:37:05, 12.84s/it]2026-04-12 14:18:36,324 | INFO | train_grpo_train | metrics_logged mode=train step=883

                                                      
{'loss': -0.0793, 'grad_norm': 1.5109801292419434, 'learning_rate': 7.3272727272727285e-06, 'num_tokens': 1799484.0, 'completions/mean_length': 220.75, 'completions/min_length': 43.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 46.0, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.7229549288749695, 'rewards/meter/std': 0.3531128466129303, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.9609307050704956, 'rewards/lexical_plausibility/std': 0.05956076458096504, 'rewards/judge_quality/mean': 0.33375000953674316, 'rewards/judge_quality/std': 0.29587340354919434, 'rewards/repeat_penalty/mean': 0.8986272811889648, 'rewards/repeat_penalty/std': 0.11397247761487961, 'rewards/near_duplicate_penalty/mean': 0.940723180770874, 'rewards/near_duplicate_penalty/std': 0.043794725090265274, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.953309178352356, 'rewards/distinct_2/std': 0.09601422399282455, 'rewards/total_composite/mean': 0.29068809747695923, 'rewards/total_composite/std': 0.2798406481742859, 'reward': 0.29068809747695923, 'reward_std': 0.2798406481742859, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11197210103273392, 'sampling/sampling_logp_difference/max': 1.4107227325439453, 'sampling/importance_sampling_ratio/min': 0.24396690726280212, 'sampling/importance_sampling_ratio/mean': 0.9942339062690735, 'sampling/importance_sampling_ratio/max': 1.7780050039291382, 'entropy': 0.4382961876690388, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.06563917454332113, 'clip_ratio/high_max': 0.06563917454332113, 'clip_ratio/region_mean': 0.06563917454332113, 'reward_total_mean': 0.29068809747695923, 'reward_meter_mean': 0.7229549288749695, 'reward_meter_std': 0.3531128466129303, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.9609307050704956, 'reward_lexical_plausibility_std': 0.05956076458096504, 'reward_repeat_penalty_mean': 0.8986272811889648, 'reward_repeat_penalty_std': 0.11397247761487961, 'reward_near_duplicate_penalty_mean': 0.940723180770874, 'reward_near_duplicate_penalty_std': 0.043794725090265274, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.953309178352356, 'reward_distinct_2_std': 0.09601422399282455, 'reward_judge_quality_mean': 0.33375000953674316, 'reward_judge_quality_std': 0.29587340354919434, 'reward_total_composite_mean': 0.29068809747695923, 'reward_total_composite_std': 0.2798406481742859, 'epoch': 0.04}

 27%|██▋       | 883/3300 [3:32:24<8:37:05, 12.84s/it]INFO 04-12 14:18:36 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 884/3300 [3:32:34<8:02:55, 11.99s/it]2026-04-12 14:18:46,351 | INFO | train_grpo_train | metrics_logged mode=train step=884

                                                      
{'loss': 0.0247, 'grad_norm': 8.392919540405273, 'learning_rate': 7.324242424242425e-06, 'num_tokens': 1802612.0, 'completions/mean_length': 182.0, 'completions/min_length': 159.0, 'completions/max_length': 213.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 182.0, 'completions/min_terminated_length': 159.0, 'completions/max_terminated_length': 213.0, 'rewards/meter/mean': 0.6464889645576477, 'rewards/meter/std': 0.21471653878688812, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.2875000238418579, 'rewards/judge_quality/std': 0.09161253273487091, 'rewards/repeat_penalty/mean': 0.6105477809906006, 'rewards/repeat_penalty/std': 0.22797583043575287, 'rewards/near_duplicate_penalty/mean': 0.8850646018981934, 'rewards/near_duplicate_penalty/std': 0.11498953402042389, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.6659486293792725, 'rewards/distinct_2/std': 0.2227121740579605, 'rewards/total_composite/mean': 0.1691592037677765, 'rewards/total_composite/std': 0.10603190213441849, 'reward': 0.1691592037677765, 'reward_std': 0.10603189468383789, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10739102214574814, 'sampling/sampling_logp_difference/max': 2.722109079360962, 'sampling/importance_sampling_ratio/min': 0.06573596596717834, 'sampling/importance_sampling_ratio/mean': 0.9946879744529724, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5474977977573872, 'clip_ratio/low_mean': 0.047079429030418396, 'clip_ratio/low_min': 0.047079429030418396, 'clip_ratio/high_mean': 0.0636939350515604, 'clip_ratio/high_max': 0.0636939350515604, 'clip_ratio/region_mean': 0.1107733640819788, 'reward_total_mean': 0.1691592037677765, 'reward_meter_mean': 0.6464889645576477, 'reward_meter_std': 0.21471653878688812, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6105477809906006, 'reward_repeat_penalty_std': 0.22797583043575287, 'reward_near_duplicate_penalty_mean': 0.8850646018981934, 'reward_near_duplicate_penalty_std': 0.11498953402042389, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.6659486293792725, 'reward_distinct_2_std': 0.2227121740579605, 'reward_judge_quality_mean': 0.2875000238418579, 'reward_judge_quality_std': 0.09161253273487091, 'reward_total_composite_mean': 0.1691592037677765, 'reward_total_composite_std': 0.10603190213441849, 'epoch': 0.04}

 27%|██▋       | 884/3300 [3:32:34<8:02:55, 11.99s/it]INFO 04-12 14:18:46 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 885/3300 [3:32:42<7:15:01, 10.81s/it]2026-04-12 14:18:54,399 | INFO | train_grpo_train | metrics_logged mode=train step=885

                                                      
{'loss': -0.0351, 'grad_norm': 10.396985054016113, 'learning_rate': 7.321212121212122e-06, 'num_tokens': 1804290.0, 'completions/mean_length': 47.75, 'completions/min_length': 26.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.75, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.7428503036499023, 'rewards/meter/std': 0.31914523243904114, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5587500333786011, 'rewards/judge_quality/std': 0.195041224360466, 'rewards/repeat_penalty/mean': 0.7894636988639832, 'rewards/repeat_penalty/std': 0.0972859337925911, 'rewards/near_duplicate_penalty/mean': 0.9249110221862793, 'rewards/near_duplicate_penalty/std': 0.05529322847723961, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.881542980670929, 'rewards/distinct_2/std': 0.05211632698774338, 'rewards/total_composite/mean': 0.4224734902381897, 'rewards/total_composite/std': 0.2751162350177765, 'reward': 0.4224734902381897, 'reward_std': 0.2751162350177765, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09754500538110733, 'sampling/sampling_logp_difference/max': 1.0492095947265625, 'sampling/importance_sampling_ratio/min': 0.35115674138069153, 'sampling/importance_sampling_ratio/mean': 1.0087987184524536, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4358046241104603, 'clip_ratio/low_mean': 0.061739190481603146, 'clip_ratio/low_min': 0.061739190481603146, 'clip_ratio/high_mean': 0.03851819969713688, 'clip_ratio/high_max': 0.03851819969713688, 'clip_ratio/region_mean': 0.10025739017874002, 'reward_total_mean': 0.4224734902381897, 'reward_meter_mean': 0.7428503036499023, 'reward_meter_std': 0.31914523243904114, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7894636988639832, 'reward_repeat_penalty_std': 0.0972859337925911, 'reward_near_duplicate_penalty_mean': 0.9249110221862793, 'reward_near_duplicate_penalty_std': 0.05529322847723961, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.881542980670929, 'reward_distinct_2_std': 0.05211632698774338, 'reward_judge_quality_mean': 0.5587500333786011, 'reward_judge_quality_std': 0.195041224360466, 'reward_total_composite_mean': 0.4224734902381897, 'reward_total_composite_std': 0.2751162350177765, 'epoch': 0.04}

 27%|██▋       | 885/3300 [3:32:42<7:15:01, 10.81s/it]INFO 04-12 14:18:54 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 886/3300 [3:32:55<7:39:05, 11.41s/it]2026-04-12 14:19:07,208 | INFO | train_grpo_train | metrics_logged mode=train step=886

                                                      
{'loss': 0.05, 'grad_norm': 5.040018558502197, 'learning_rate': 7.3181818181818186e-06, 'num_tokens': 1808849.0, 'completions/mean_length': 330.875, 'completions/min_length': 310.0, 'completions/max_length': 382.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 330.875, 'completions/min_terminated_length': 310.0, 'completions/max_terminated_length': 382.0, 'rewards/meter/mean': 0.09013642370700836, 'rewards/meter/std': 0.08543291687965393, 'rewards/count_adherence/mean': 0.8250000476837158, 'rewards/count_adherence/std': 0.0707106664776802, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.2519568204879761, 'rewards/repeat_penalty/std': 0.1718509942293167, 'rewards/near_duplicate_penalty/mean': 0.740941047668457, 'rewards/near_duplicate_penalty/std': 0.09234555810689926, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.3602747619152069, 'rewards/distinct_2/std': 0.210606649518013, 'rewards/total_composite/mean': 0.01846323534846306, 'rewards/total_composite/std': 0.01902730017900467, 'reward': 0.01846323534846306, 'reward_std': 0.01902730017900467, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0903920903801918, 'sampling/sampling_logp_difference/max': 2.914473533630371, 'sampling/importance_sampling_ratio/min': 0.05423257499933243, 'sampling/importance_sampling_ratio/mean': 1.0015575885772705, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.39976517111063004, 'clip_ratio/low_mean': 0.04512396873906255, 'clip_ratio/low_min': 0.04512396873906255, 'clip_ratio/high_mean': 0.035638079047203064, 'clip_ratio/high_max': 0.035638079047203064, 'clip_ratio/region_mean': 0.08076204778626561, 'reward_total_mean': 0.01846323534846306, 'reward_meter_mean': 0.09013642370700836, 'reward_meter_std': 0.08543291687965393, 'reward_count_adherence_mean': 0.8250000476837158, 'reward_count_adherence_std': 0.0707106664776802, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.2519568204879761, 'reward_repeat_penalty_std': 0.1718509942293167, 'reward_near_duplicate_penalty_mean': 0.740941047668457, 'reward_near_duplicate_penalty_std': 0.09234555810689926, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.3602747619152069, 'reward_distinct_2_std': 0.210606649518013, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.01846323534846306, 'reward_total_composite_std': 0.01902730017900467, 'epoch': 0.04}

 27%|██▋       | 886/3300 [3:32:55<7:39:05, 11.41s/it]INFO 04-12 14:19:07 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 887/3300 [3:33:08<7:59:06, 11.91s/it]2026-04-12 14:19:20,290 | INFO | train_grpo_train | metrics_logged mode=train step=887

                                                      
{'loss': -0.0799, 'grad_norm': 4.748170375823975, 'learning_rate': 7.315151515151516e-06, 'num_tokens': 1810698.0, 'completions/mean_length': 130.125, 'completions/min_length': 68.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 75.5714340209961, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.5883125066757202, 'rewards/meter/std': 0.42992734909057617, '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/lexical_plausibility/mean': 0.9682332873344421, 'rewards/lexical_plausibility/std': 0.08984983712434769, 'rewards/judge_quality/mean': 0.4300000071525574, 'rewards/judge_quality/std': 0.32262319326400757, 'rewards/repeat_penalty/mean': 0.6742799282073975, 'rewards/repeat_penalty/std': 0.16807249188423157, 'rewards/near_duplicate_penalty/mean': 0.9275095462799072, 'rewards/near_duplicate_penalty/std': 0.03420286625623703, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.8833611011505127, 'rewards/distinct_2/std': 0.083704374730587, 'rewards/total_composite/mean': 0.24122354388237, 'rewards/total_composite/std': 0.1925996094942093, 'reward': 0.24122354388237, 'reward_std': 0.1925996094942093, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10349947214126587, 'sampling/sampling_logp_difference/max': 2.3311681747436523, 'sampling/importance_sampling_ratio/min': 0.09718216210603714, 'sampling/importance_sampling_ratio/mean': 1.0020560026168823, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.41072361543774605, 'clip_ratio/low_mean': 0.03469423530623317, 'clip_ratio/low_min': 0.03469423530623317, 'clip_ratio/high_mean': 0.04431463126093149, 'clip_ratio/high_max': 0.04431463126093149, 'clip_ratio/region_mean': 0.07900886656716466, 'reward_total_mean': 0.24122354388237, 'reward_meter_mean': 0.5883125066757202, 'reward_meter_std': 0.42992734909057617, '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_lexical_plausibility_mean': 0.9682332873344421, 'reward_lexical_plausibility_std': 0.08984983712434769, 'reward_repeat_penalty_mean': 0.6742799282073975, 'reward_repeat_penalty_std': 0.16807249188423157, 'reward_near_duplicate_penalty_mean': 0.9275095462799072, 'reward_near_duplicate_penalty_std': 0.03420286625623703, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.8833611011505127, 'reward_distinct_2_std': 0.083704374730587, 'reward_judge_quality_mean': 0.4300000071525574, 'reward_judge_quality_std': 0.32262319326400757, 'reward_total_composite_mean': 0.24122354388237, 'reward_total_composite_std': 0.1925996094942093, 'epoch': 0.04}

 27%|██▋       | 887/3300 [3:33:08<7:59:06, 11.91s/it]INFO 04-12 14:19:20 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 888/3300 [3:33:21<8:19:54, 12.44s/it]2026-04-12 14:19:33,955 | INFO | train_grpo_train | metrics_logged mode=train step=888

                                                      
{'loss': -0.029, 'grad_norm': 6.361115455627441, 'learning_rate': 7.312121212121212e-06, 'num_tokens': 1812533.0, 'completions/mean_length': 115.375, 'completions/min_length': 53.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 58.71428680419922, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.24721597135066986, 'rewards/meter/std': 0.30799585580825806, '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/lexical_plausibility/mean': 0.9663930535316467, 'rewards/lexical_plausibility/std': 0.09505482017993927, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.8729970455169678, 'rewards/repeat_penalty/std': 0.0752529501914978, 'rewards/near_duplicate_penalty/mean': 0.9179536700248718, 'rewards/near_duplicate_penalty/std': 0.025116555392742157, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9506443738937378, 'rewards/distinct_2/std': 0.0722939670085907, 'rewards/total_composite/mean': 0.09125371277332306, 'rewards/total_composite/std': 0.10675454139709473, 'reward': 0.09125371277332306, 'reward_std': 0.10675454139709473, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13567937910556793, 'sampling/sampling_logp_difference/max': 2.0671396255493164, 'sampling/importance_sampling_ratio/min': 0.12654723227024078, 'sampling/importance_sampling_ratio/mean': 1.011005163192749, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5403723903000355, 'clip_ratio/low_mean': 0.05108470655977726, 'clip_ratio/low_min': 0.05108470655977726, 'clip_ratio/high_mean': 0.04891751799732447, 'clip_ratio/high_max': 0.04891751799732447, 'clip_ratio/region_mean': 0.10000222455710173, 'reward_total_mean': 0.09125371277332306, 'reward_meter_mean': 0.24721597135066986, 'reward_meter_std': 0.30799585580825806, '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_lexical_plausibility_mean': 0.9663930535316467, 'reward_lexical_plausibility_std': 0.09505482017993927, 'reward_repeat_penalty_mean': 0.8729970455169678, 'reward_repeat_penalty_std': 0.0752529501914978, 'reward_near_duplicate_penalty_mean': 0.9179536700248718, 'reward_near_duplicate_penalty_std': 0.025116555392742157, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9506443738937378, 'reward_distinct_2_std': 0.0722939670085907, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.09125371277332306, 'reward_total_composite_std': 0.10675454139709473, 'epoch': 0.04}

 27%|██▋       | 888/3300 [3:33:21<8:19:54, 12.44s/it]INFO 04-12 14:19:34 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 889/3300 [3:33:35<8:31:55, 12.74s/it]2026-04-12 14:19:47,402 | INFO | train_grpo_train | metrics_logged mode=train step=889

                                                      
{'loss': -0.0597, 'grad_norm': 4.837181568145752, 'learning_rate': 7.30909090909091e-06, 'num_tokens': 1814164.0, 'completions/mean_length': 109.875, 'completions/min_length': 48.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 52.42857360839844, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.5529102683067322, 'rewards/meter/std': 0.4197533428668976, '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/lexical_plausibility/mean': 0.9508507251739502, 'rewards/lexical_plausibility/std': 0.13901503384113312, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.6603304147720337, 'rewards/repeat_penalty/std': 0.311467707157135, 'rewards/near_duplicate_penalty/mean': 0.8855952024459839, 'rewards/near_duplicate_penalty/std': 0.08292616158723831, 'rewards/opening_diversity/mean': 0.828125, 'rewards/opening_diversity/std': 0.22097088396549225, 'rewards/distinct_2/mean': 0.8661375641822815, 'rewards/distinct_2/std': 0.16727451980113983, 'rewards/total_composite/mean': 0.1800411343574524, 'rewards/total_composite/std': 0.15511508285999298, 'reward': 0.1800411343574524, 'reward_std': 0.15511509776115417, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0980653464794159, 'sampling/sampling_logp_difference/max': 1.3906984329223633, 'sampling/importance_sampling_ratio/min': 0.24890141189098358, 'sampling/importance_sampling_ratio/mean': 1.011983036994934, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4984290860593319, 'clip_ratio/low_mean': 0.027500967727974057, 'clip_ratio/low_min': 0.027500967727974057, 'clip_ratio/high_mean': 0.041873543756082654, 'clip_ratio/high_max': 0.041873543756082654, 'clip_ratio/region_mean': 0.06937451148405671, 'reward_total_mean': 0.1800411343574524, 'reward_meter_mean': 0.5529102683067322, 'reward_meter_std': 0.4197533428668976, '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_lexical_plausibility_mean': 0.9508507251739502, 'reward_lexical_plausibility_std': 0.13901503384113312, 'reward_repeat_penalty_mean': 0.6603304147720337, 'reward_repeat_penalty_std': 0.311467707157135, 'reward_near_duplicate_penalty_mean': 0.8855952024459839, 'reward_near_duplicate_penalty_std': 0.08292616158723831, 'reward_opening_diversity_mean': 0.828125, 'reward_opening_diversity_std': 0.22097088396549225, 'reward_distinct_2_mean': 0.8661375641822815, 'reward_distinct_2_std': 0.16727451980113983, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.1800411343574524, 'reward_total_composite_std': 0.15511508285999298, 'epoch': 0.04}

 27%|██▋       | 889/3300 [3:33:35<8:31:55, 12.74s/it]INFO 04-12 14:19:47 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 890/3300 [3:33:44<7:46:47, 11.62s/it]2026-04-12 14:19:56,413 | INFO | train_grpo_train | metrics_logged mode=train step=890

                                                      
{'loss': 0.0958, 'grad_norm': 8.40279769897461, 'learning_rate': 7.306060606060607e-06, 'num_tokens': 1816104.0, 'completions/mean_length': 89.5, 'completions/min_length': 62.0, 'completions/max_length': 118.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 89.5, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 118.0, 'rewards/meter/mean': 0.48707133531570435, 'rewards/meter/std': 0.29376834630966187, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.17500001192092896, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/repeat_penalty/mean': 0.31250518560409546, 'rewards/repeat_penalty/std': 0.29178664088249207, 'rewards/near_duplicate_penalty/mean': 0.7215415239334106, 'rewards/near_duplicate_penalty/std': 0.2148185670375824, 'rewards/opening_diversity/mean': 0.7942708730697632, 'rewards/opening_diversity/std': 0.28225407004356384, 'rewards/distinct_2/mean': 0.40195611119270325, 'rewards/distinct_2/std': 0.3049626052379608, 'rewards/total_composite/mean': 0.08777297288179398, 'rewards/total_composite/std': 0.06469108909368515, 'reward': 0.08777297288179398, 'reward_std': 0.06469108909368515, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09963199496269226, 'sampling/sampling_logp_difference/max': 2.1541762351989746, 'sampling/importance_sampling_ratio/min': 0.11599870771169662, 'sampling/importance_sampling_ratio/mean': 1.0099962949752808, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5191951468586922, 'clip_ratio/low_mean': 0.04910414153710008, 'clip_ratio/low_min': 0.04910414153710008, 'clip_ratio/high_mean': 0.04114343272522092, 'clip_ratio/high_max': 0.04114343272522092, 'clip_ratio/region_mean': 0.090247574262321, 'reward_total_mean': 0.08777297288179398, 'reward_meter_mean': 0.48707133531570435, 'reward_meter_std': 0.29376834630966187, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.31250518560409546, 'reward_repeat_penalty_std': 0.29178664088249207, 'reward_near_duplicate_penalty_mean': 0.7215415239334106, 'reward_near_duplicate_penalty_std': 0.2148185670375824, 'reward_opening_diversity_mean': 0.7942708730697632, 'reward_opening_diversity_std': 0.28225407004356384, 'reward_distinct_2_mean': 0.40195611119270325, 'reward_distinct_2_std': 0.3049626052379608, 'reward_judge_quality_mean': 0.17500001192092896, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.08777297288179398, 'reward_total_composite_std': 0.06469108909368515, 'epoch': 0.04}

 27%|██▋       | 890/3300 [3:33:44<7:46:47, 11.62s/it]INFO 04-12 14:19:56 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 891/3300 [3:33:57<8:03:54, 12.05s/it]2026-04-12 14:20:09,467 | INFO | train_grpo_train | metrics_logged mode=train step=891

                                                      
{'loss': -0.0295, 'grad_norm': 1.2182680368423462, 'learning_rate': 7.303030303030304e-06, 'num_tokens': 1817386.0, 'completions/mean_length': 333.25, 'completions/min_length': 31.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.625, 'completions/mean_terminated_length': 35.333335876464844, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.28022828698158264, 'rewards/meter/std': 0.34886088967323303, '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/lexical_plausibility/mean': 0.8258205056190491, 'rewards/lexical_plausibility/std': 0.1288675218820572, 'rewards/judge_quality/mean': 0.2150000035762787, 'rewards/judge_quality/std': 0.3188372254371643, 'rewards/repeat_penalty/mean': 0.9227318167686462, 'rewards/repeat_penalty/std': 0.09093921631574631, 'rewards/near_duplicate_penalty/mean': 0.9550496935844421, 'rewards/near_duplicate_penalty/std': 0.0405331626534462, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9648663997650146, 'rewards/distinct_2/std': 0.06941013783216476, 'rewards/total_composite/mean': 0.10385888814926147, 'rewards/total_composite/std': 0.1914631426334381, 'reward': 0.10385888814926147, 'reward_std': 0.1914631426334381, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1345459371805191, 'sampling/sampling_logp_difference/max': 1.0366363525390625, 'sampling/importance_sampling_ratio/min': 0.3914050757884979, 'sampling/importance_sampling_ratio/mean': 1.0274628400802612, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3371424302458763, 'clip_ratio/low_mean': 0.0029761905316263437, 'clip_ratio/low_min': 0.0029761905316263437, 'clip_ratio/high_mean': 0.04337732121348381, 'clip_ratio/high_max': 0.04337732121348381, 'clip_ratio/region_mean': 0.046353511745110154, 'reward_total_mean': 0.10385888814926147, 'reward_meter_mean': 0.28022828698158264, 'reward_meter_std': 0.34886088967323303, '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_lexical_plausibility_mean': 0.8258205056190491, 'reward_lexical_plausibility_std': 0.1288675218820572, 'reward_repeat_penalty_mean': 0.9227318167686462, 'reward_repeat_penalty_std': 0.09093921631574631, 'reward_near_duplicate_penalty_mean': 0.9550496935844421, 'reward_near_duplicate_penalty_std': 0.0405331626534462, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9648663997650146, 'reward_distinct_2_std': 0.06941013783216476, 'reward_judge_quality_mean': 0.2150000035762787, 'reward_judge_quality_std': 0.3188372254371643, 'reward_total_composite_mean': 0.10385888814926147, 'reward_total_composite_std': 0.1914631426334381, 'epoch': 0.04}

 27%|██▋       | 891/3300 [3:33:57<8:03:54, 12.05s/it]INFO 04-12 14:20:09 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 892/3300 [3:34:11<8:24:43, 12.58s/it]2026-04-12 14:20:23,269 | INFO | train_grpo_train | metrics_logged mode=train step=892

                                                      
{'loss': -0.0432, 'grad_norm': 5.134486675262451, 'learning_rate': 7.3e-06, 'num_tokens': 1819190.0, 'completions/mean_length': 116.5, 'completions/min_length': 53.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 60.000003814697266, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.7746455073356628, 'rewards/meter/std': 0.3026490807533264, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9716020822525024, 'rewards/lexical_plausibility/std': 0.08032133430242538, 'rewards/judge_quality/mean': 0.39625000953674316, 'rewards/judge_quality/std': 0.2585087716579437, 'rewards/repeat_penalty/mean': 0.9318588972091675, 'rewards/repeat_penalty/std': 0.0886862576007843, 'rewards/near_duplicate_penalty/mean': 0.9741957187652588, 'rewards/near_duplicate_penalty/std': 0.02595910243690014, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9553326368331909, 'rewards/distinct_2/std': 0.07399500906467438, 'rewards/total_composite/mean': 0.2802067995071411, 'rewards/total_composite/std': 0.306781530380249, 'reward': 0.2802067995071411, 'reward_std': 0.306781530380249, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12372325360774994, 'sampling/sampling_logp_difference/max': 1.265763282775879, 'sampling/importance_sampling_ratio/min': 0.28202393651008606, 'sampling/importance_sampling_ratio/mean': 1.0104753971099854, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6737786903977394, 'clip_ratio/low_mean': 0.0495247570797801, 'clip_ratio/low_min': 0.0495247570797801, 'clip_ratio/high_mean': 0.06382819265127182, 'clip_ratio/high_max': 0.06382819265127182, 'clip_ratio/region_mean': 0.11335294973105192, 'reward_total_mean': 0.2802067995071411, 'reward_meter_mean': 0.7746455073356628, 'reward_meter_std': 0.3026490807533264, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9716020822525024, 'reward_lexical_plausibility_std': 0.08032133430242538, 'reward_repeat_penalty_mean': 0.9318588972091675, 'reward_repeat_penalty_std': 0.0886862576007843, 'reward_near_duplicate_penalty_mean': 0.9741957187652588, 'reward_near_duplicate_penalty_std': 0.02595910243690014, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9553326368331909, 'reward_distinct_2_std': 0.07399500906467438, 'reward_judge_quality_mean': 0.39625000953674316, 'reward_judge_quality_std': 0.2585087716579437, 'reward_total_composite_mean': 0.2802067995071411, 'reward_total_composite_std': 0.306781530380249, 'epoch': 0.04}

 27%|██▋       | 892/3300 [3:34:11<8:24:43, 12.58s/it]INFO 04-12 14:20:23 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 893/3300 [3:34:24<8:38:41, 12.93s/it]2026-04-12 14:20:37,025 | INFO | train_grpo_train | metrics_logged mode=train step=893

                                                      
{'loss': -0.1405, 'grad_norm': 3.605888843536377, 'learning_rate': 7.296969696969698e-06, 'num_tokens': 1821401.0, 'completions/mean_length': 213.375, 'completions/min_length': 106.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 113.83333587646484, 'completions/min_terminated_length': 106.0, 'completions/max_terminated_length': 117.0, 'rewards/meter/mean': 0.5278173685073853, 'rewards/meter/std': 0.380450576543808, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9403489828109741, 'rewards/lexical_plausibility/std': 0.11958342045545578, 'rewards/judge_quality/mean': 0.28125, 'rewards/judge_quality/std': 0.13346347212791443, 'rewards/repeat_penalty/mean': 0.7489525079727173, 'rewards/repeat_penalty/std': 0.16229791939258575, 'rewards/near_duplicate_penalty/mean': 0.9234349727630615, 'rewards/near_duplicate_penalty/std': 0.04624555632472038, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8060756325721741, 'rewards/distinct_2/std': 0.14574405550956726, 'rewards/total_composite/mean': 0.13932116329669952, 'rewards/total_composite/std': 0.1388460397720337, 'reward': 0.13932116329669952, 'reward_std': 0.1388460397720337, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12520933151245117, 'sampling/sampling_logp_difference/max': 3.3137621879577637, 'sampling/importance_sampling_ratio/min': 0.036379050463438034, 'sampling/importance_sampling_ratio/mean': 1.0155220031738281, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5951655730605125, 'clip_ratio/low_mean': 0.038522012531757355, 'clip_ratio/low_min': 0.038522012531757355, 'clip_ratio/high_mean': 0.06053047347813845, 'clip_ratio/high_max': 0.06053047347813845, 'clip_ratio/region_mean': 0.0990524860098958, 'reward_total_mean': 0.13932116329669952, 'reward_meter_mean': 0.5278173685073853, 'reward_meter_std': 0.380450576543808, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9403489828109741, 'reward_lexical_plausibility_std': 0.11958342045545578, 'reward_repeat_penalty_mean': 0.7489525079727173, 'reward_repeat_penalty_std': 0.16229791939258575, 'reward_near_duplicate_penalty_mean': 0.9234349727630615, 'reward_near_duplicate_penalty_std': 0.04624555632472038, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8060756325721741, 'reward_distinct_2_std': 0.14574405550956726, 'reward_judge_quality_mean': 0.28125, 'reward_judge_quality_std': 0.13346347212791443, 'reward_total_composite_mean': 0.13932116329669952, 'reward_total_composite_std': 0.1388460397720337, 'epoch': 0.04}

 27%|██▋       | 893/3300 [3:34:24<8:38:41, 12.93s/it]INFO 04-12 14:20:37 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 894/3300 [3:34:39<8:53:32, 13.31s/it]2026-04-12 14:20:51,205 | INFO | train_grpo_train | metrics_logged mode=train step=894

                                                      
{'loss': -0.0537, 'grad_norm': 4.7556352615356445, 'learning_rate': 7.293939393939394e-06, 'num_tokens': 1823099.0, 'completions/mean_length': 109.25, 'completions/min_length': 47.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 51.71428680419922, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.6064315438270569, 'rewards/meter/std': 0.4437050223350525, '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/lexical_plausibility/mean': 0.9821075797080994, 'rewards/lexical_plausibility/std': 0.05060745030641556, 'rewards/judge_quality/mean': 0.5762500166893005, 'rewards/judge_quality/std': 0.31513887643814087, 'rewards/repeat_penalty/mean': 0.8857088685035706, 'rewards/repeat_penalty/std': 0.06341800838708878, 'rewards/near_duplicate_penalty/mean': 0.9281342625617981, 'rewards/near_duplicate_penalty/std': 0.051846202462911606, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9541317820549011, 'rewards/distinct_2/std': 0.03692275285720825, 'rewards/total_composite/mean': 0.3359679579734802, 'rewards/total_composite/std': 0.3716164231300354, 'reward': 0.3359679579734802, 'reward_std': 0.371616393327713, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09344349056482315, 'sampling/sampling_logp_difference/max': 1.2455737590789795, 'sampling/importance_sampling_ratio/min': 0.28777575492858887, 'sampling/importance_sampling_ratio/mean': 1.0080801248550415, 'sampling/importance_sampling_ratio/max': 1.9176734685897827, 'entropy': 0.443700160831213, 'clip_ratio/low_mean': 0.0349231674335897, 'clip_ratio/low_min': 0.0349231674335897, 'clip_ratio/high_mean': 0.03835763339884579, 'clip_ratio/high_max': 0.03835763339884579, 'clip_ratio/region_mean': 0.07328080083243549, 'reward_total_mean': 0.3359679579734802, 'reward_meter_mean': 0.6064315438270569, 'reward_meter_std': 0.4437050223350525, '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_lexical_plausibility_mean': 0.9821075797080994, 'reward_lexical_plausibility_std': 0.05060745030641556, 'reward_repeat_penalty_mean': 0.8857088685035706, 'reward_repeat_penalty_std': 0.06341800838708878, 'reward_near_duplicate_penalty_mean': 0.9281342625617981, 'reward_near_duplicate_penalty_std': 0.051846202462911606, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9541317820549011, 'reward_distinct_2_std': 0.03692275285720825, 'reward_judge_quality_mean': 0.5762500166893005, 'reward_judge_quality_std': 0.31513887643814087, 'reward_total_composite_mean': 0.3359679579734802, 'reward_total_composite_std': 0.3716164231300354, 'epoch': 0.04}

 27%|██▋       | 894/3300 [3:34:39<8:53:32, 13.31s/it]INFO 04-12 14:20:51 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 895/3300 [3:34:52<8:59:07, 13.45s/it]2026-04-12 14:21:04,996 | INFO | train_grpo_train | metrics_logged mode=train step=895

                                                      
{'loss': -0.0466, 'grad_norm': 2.282470464706421, 'learning_rate': 7.290909090909092e-06, 'num_tokens': 1824720.0, 'completions/mean_length': 226.625, 'completions/min_length': 50.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 55.400001525878906, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.24938635528087616, 'rewards/meter/std': 0.3241385519504547, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.8015584945678711, 'rewards/lexical_plausibility/std': 0.12472990155220032, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.16690459847450256, 'rewards/repeat_penalty/mean': 0.7836134433746338, 'rewards/repeat_penalty/std': 0.14723502099514008, 'rewards/near_duplicate_penalty/mean': 0.9140781164169312, 'rewards/near_duplicate_penalty/std': 0.0649443045258522, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9146628379821777, 'rewards/distinct_2/std': 0.10737480968236923, 'rewards/total_composite/mean': 0.09016980975866318, 'rewards/total_composite/std': 0.14305296540260315, 'reward': 0.09016980975866318, 'reward_std': 0.14305298030376434, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1548057645559311, 'sampling/sampling_logp_difference/max': 1.9008073806762695, 'sampling/importance_sampling_ratio/min': 0.14944791793823242, 'sampling/importance_sampling_ratio/mean': 1.032528042793274, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7295360565185547, 'clip_ratio/low_mean': 0.020233366638422012, 'clip_ratio/low_min': 0.020233366638422012, 'clip_ratio/high_mean': 0.04418042488396168, 'clip_ratio/high_max': 0.04418042488396168, 'clip_ratio/region_mean': 0.06441379152238369, 'reward_total_mean': 0.09016980975866318, 'reward_meter_mean': 0.24938635528087616, 'reward_meter_std': 0.3241385519504547, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.8015584945678711, 'reward_lexical_plausibility_std': 0.12472990155220032, 'reward_repeat_penalty_mean': 0.7836134433746338, 'reward_repeat_penalty_std': 0.14723502099514008, 'reward_near_duplicate_penalty_mean': 0.9140781164169312, 'reward_near_duplicate_penalty_std': 0.0649443045258522, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9146628379821777, 'reward_distinct_2_std': 0.10737480968236923, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.16690459847450256, 'reward_total_composite_mean': 0.09016980975866318, 'reward_total_composite_std': 0.14305296540260315, 'epoch': 0.04}

 27%|██▋       | 895/3300 [3:34:52<8:59:07, 13.45s/it]INFO 04-12 14:21:05 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 896/3300 [3:35:06<8:59:48, 13.47s/it]2026-04-12 14:21:18,523 | INFO | train_grpo_train | metrics_logged mode=train step=896

                                                      
{'loss': -0.0902, 'grad_norm': 1.4859238862991333, 'learning_rate': 7.287878787878789e-06, 'num_tokens': 1826222.0, 'completions/mean_length': 223.75, 'completions/min_length': 45.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 50.79999923706055, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.571113109588623, 'rewards/meter/std': 0.35994669795036316, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.911432683467865, 'rewards/lexical_plausibility/std': 0.13451498746871948, 'rewards/judge_quality/mean': 0.3087500035762787, 'rewards/judge_quality/std': 0.2930839955806732, 'rewards/repeat_penalty/mean': 0.9230127334594727, 'rewards/repeat_penalty/std': 0.09052017331123352, 'rewards/near_duplicate_penalty/mean': 0.9542627334594727, 'rewards/near_duplicate_penalty/std': 0.06040158122777939, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2108093649148941, 'rewards/total_composite/std': 0.20128123462200165, 'reward': 0.2108093649148941, 'reward_std': 0.20128123462200165, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12994879484176636, 'sampling/sampling_logp_difference/max': 1.7714948654174805, 'sampling/importance_sampling_ratio/min': 0.1700785607099533, 'sampling/importance_sampling_ratio/mean': 1.0071216821670532, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4746543578803539, 'clip_ratio/low_mean': 0.01944444514811039, 'clip_ratio/low_min': 0.01944444514811039, 'clip_ratio/high_mean': 0.052552847657352686, 'clip_ratio/high_max': 0.052552847657352686, 'clip_ratio/region_mean': 0.07199729280546308, 'reward_total_mean': 0.2108093649148941, 'reward_meter_mean': 0.571113109588623, 'reward_meter_std': 0.35994669795036316, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.911432683467865, 'reward_lexical_plausibility_std': 0.13451498746871948, 'reward_repeat_penalty_mean': 0.9230127334594727, 'reward_repeat_penalty_std': 0.09052017331123352, 'reward_near_duplicate_penalty_mean': 0.9542627334594727, 'reward_near_duplicate_penalty_std': 0.06040158122777939, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3087500035762787, 'reward_judge_quality_std': 0.2930839955806732, 'reward_total_composite_mean': 0.2108093649148941, 'reward_total_composite_std': 0.20128123462200165, 'epoch': 0.04}

 27%|██▋       | 896/3300 [3:35:06<8:59:48, 13.47s/it]INFO 04-12 14:21:18 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 897/3300 [3:35:21<9:13:50, 13.83s/it]2026-04-12 14:21:33,177 | INFO | train_grpo_train | metrics_logged mode=train step=897

                                                      
{'loss': -0.0846, 'grad_norm': 3.193230628967285, 'learning_rate': 7.284848484848486e-06, 'num_tokens': 1827979.0, 'completions/mean_length': 170.625, 'completions/min_length': 54.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 56.833335876464844, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.32843998074531555, 'rewards/meter/std': 0.31348317861557007, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9387256503105164, 'rewards/lexical_plausibility/std': 0.11420927196741104, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.18077214062213898, 'rewards/repeat_penalty/mean': 0.918314516544342, 'rewards/repeat_penalty/std': 0.09387185424566269, 'rewards/near_duplicate_penalty/mean': 0.9749169945716858, 'rewards/near_duplicate_penalty/std': 0.033262208104133606, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9729020595550537, 'rewards/distinct_2/std': 0.037460874766111374, 'rewards/total_composite/mean': 0.12396536767482758, 'rewards/total_composite/std': 0.12815655767917633, 'reward': 0.12396536767482758, 'reward_std': 0.12815655767917633, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11207740753889084, 'sampling/sampling_logp_difference/max': 1.2449564933776855, 'sampling/importance_sampling_ratio/min': 0.3877437710762024, 'sampling/importance_sampling_ratio/mean': 1.0083962678909302, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4012759141623974, 'clip_ratio/low_mean': 0.029596560634672642, 'clip_ratio/low_min': 0.029596560634672642, 'clip_ratio/high_mean': 0.05037395842373371, 'clip_ratio/high_max': 0.05037395842373371, 'clip_ratio/region_mean': 0.07997051905840635, 'reward_total_mean': 0.12396536767482758, 'reward_meter_mean': 0.32843998074531555, 'reward_meter_std': 0.31348317861557007, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9387256503105164, 'reward_lexical_plausibility_std': 0.11420927196741104, 'reward_repeat_penalty_mean': 0.918314516544342, 'reward_repeat_penalty_std': 0.09387185424566269, 'reward_near_duplicate_penalty_mean': 0.9749169945716858, 'reward_near_duplicate_penalty_std': 0.033262208104133606, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9729020595550537, 'reward_distinct_2_std': 0.037460874766111374, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.18077214062213898, 'reward_total_composite_mean': 0.12396536767482758, 'reward_total_composite_std': 0.12815655767917633, 'epoch': 0.04}

 27%|██▋       | 897/3300 [3:35:21<9:13:50, 13.83s/it]INFO 04-12 14:21:33 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 898/3300 [3:35:34<9:03:33, 13.58s/it]2026-04-12 14:21:46,168 | INFO | train_grpo_train | metrics_logged mode=train step=898

                                                      
{'loss': -0.0352, 'grad_norm': 4.640687465667725, 'learning_rate': 7.281818181818182e-06, 'num_tokens': 1829442.0, 'completions/mean_length': 88.875, 'completions/min_length': 23.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 28.428573608398438, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.9009531736373901, 'rewards/meter/std': 0.21193014085292816, '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/lexical_plausibility/mean': 0.9477071762084961, 'rewards/lexical_plausibility/std': 0.099457286298275, 'rewards/judge_quality/mean': 0.2837499976158142, 'rewards/judge_quality/std': 0.27769139409065247, 'rewards/repeat_penalty/mean': 0.7404162883758545, 'rewards/repeat_penalty/std': 0.08220633864402771, 'rewards/near_duplicate_penalty/mean': 0.9521687626838684, 'rewards/near_duplicate_penalty/std': 0.09419640153646469, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9990137815475464, 'rewards/distinct_2/std': 0.002789383754134178, 'rewards/total_composite/mean': 0.24241797626018524, 'rewards/total_composite/std': 0.2762780785560608, 'reward': 0.24241797626018524, 'reward_std': 0.2762780785560608, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12997201085090637, 'sampling/sampling_logp_difference/max': 1.2856168746948242, 'sampling/importance_sampling_ratio/min': 0.2764799892902374, 'sampling/importance_sampling_ratio/mean': 0.9943518042564392, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5958964861929417, 'clip_ratio/low_mean': 0.07402955833822489, 'clip_ratio/low_min': 0.07402955833822489, 'clip_ratio/high_mean': 0.04330294393002987, 'clip_ratio/high_max': 0.04330294393002987, 'clip_ratio/region_mean': 0.11733250226825476, 'reward_total_mean': 0.24241797626018524, 'reward_meter_mean': 0.9009531736373901, 'reward_meter_std': 0.21193014085292816, '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_lexical_plausibility_mean': 0.9477071762084961, 'reward_lexical_plausibility_std': 0.099457286298275, 'reward_repeat_penalty_mean': 0.7404162883758545, 'reward_repeat_penalty_std': 0.08220633864402771, 'reward_near_duplicate_penalty_mean': 0.9521687626838684, 'reward_near_duplicate_penalty_std': 0.09419640153646469, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9990137815475464, 'reward_distinct_2_std': 0.002789383754134178, 'reward_judge_quality_mean': 0.2837499976158142, 'reward_judge_quality_std': 0.27769139409065247, 'reward_total_composite_mean': 0.24241797626018524, 'reward_total_composite_std': 0.2762780785560608, 'epoch': 0.04}

 27%|██▋       | 898/3300 [3:35:34<9:03:33, 13.58s/it]INFO 04-12 14:21:46 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 899/3300 [3:35:47<8:59:20, 13.48s/it]2026-04-12 14:21:59,414 | INFO | train_grpo_train | metrics_logged mode=train step=899

                                                      
{'loss': -0.1324, 'grad_norm': 1.8525489568710327, 'learning_rate': 7.2787878787878795e-06, 'num_tokens': 1831163.0, 'completions/mean_length': 179.125, 'completions/min_length': 60.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 68.16667175292969, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.6899092197418213, 'rewards/meter/std': 0.2848189175128937, '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/lexical_plausibility/mean': 0.9384576082229614, 'rewards/lexical_plausibility/std': 0.12220805138349533, 'rewards/judge_quality/mean': 0.39625000953674316, 'rewards/judge_quality/std': 0.27458736300468445, 'rewards/repeat_penalty/mean': 0.672481894493103, 'rewards/repeat_penalty/std': 0.2398921251296997, 'rewards/near_duplicate_penalty/mean': 0.9235548377037048, 'rewards/near_duplicate_penalty/std': 0.03125934302806854, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.24775780737400055, 'rewards/distinct_2/mean': 0.9214152097702026, 'rewards/distinct_2/std': 0.0647495687007904, 'rewards/total_composite/mean': 0.24811697006225586, 'rewards/total_composite/std': 0.1650158166885376, 'reward': 0.24811697006225586, 'reward_std': 0.1650158166885376, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12798967957496643, 'sampling/sampling_logp_difference/max': 3.1604881286621094, 'sampling/importance_sampling_ratio/min': 0.042405035346746445, 'sampling/importance_sampling_ratio/mean': 0.9993575215339661, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3327045850455761, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.08540903823450208, 'clip_ratio/high_max': 0.08540903823450208, 'clip_ratio/region_mean': 0.08540903823450208, 'reward_total_mean': 0.24811697006225586, 'reward_meter_mean': 0.6899092197418213, 'reward_meter_std': 0.2848189175128937, '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_lexical_plausibility_mean': 0.9384576082229614, 'reward_lexical_plausibility_std': 0.12220805138349533, 'reward_repeat_penalty_mean': 0.672481894493103, 'reward_repeat_penalty_std': 0.2398921251296997, 'reward_near_duplicate_penalty_mean': 0.9235548377037048, 'reward_near_duplicate_penalty_std': 0.03125934302806854, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.24775780737400055, 'reward_distinct_2_mean': 0.9214152097702026, 'reward_distinct_2_std': 0.0647495687007904, 'reward_judge_quality_mean': 0.39625000953674316, 'reward_judge_quality_std': 0.27458736300468445, 'reward_total_composite_mean': 0.24811697006225586, 'reward_total_composite_std': 0.1650158166885376, 'epoch': 0.04}

 27%|██▋       | 899/3300 [3:35:47<8:59:20, 13.48s/it]INFO 04-12 14:21:59 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 900/3300 [3:36:00<8:51:38, 13.29s/it]2026-04-12 14:22:12,273 | INFO | train_grpo_train | metrics_logged mode=train step=900

                                                      
{'loss': -0.047, 'grad_norm': 4.618988037109375, 'learning_rate': 7.275757575757576e-06, 'num_tokens': 1832644.0, 'completions/mean_length': 94.125, 'completions/min_length': 26.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 34.42857360839844, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.5552480220794678, 'rewards/meter/std': 0.35175880789756775, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9113634824752808, 'rewards/lexical_plausibility/std': 0.16448672115802765, 'rewards/judge_quality/mean': 0.5375000238418579, 'rewards/judge_quality/std': 0.3681517541408539, 'rewards/repeat_penalty/mean': 0.7949947118759155, 'rewards/repeat_penalty/std': 0.21434663236141205, 'rewards/near_duplicate_penalty/mean': 0.9128570556640625, 'rewards/near_duplicate_penalty/std': 0.06426502019166946, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.8788214921951294, 'rewards/distinct_2/std': 0.14814431965351105, 'rewards/total_composite/mean': 0.24640285968780518, 'rewards/total_composite/std': 0.20807690918445587, 'reward': 0.24640285968780518, 'reward_std': 0.20807690918445587, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12182287871837616, 'sampling/sampling_logp_difference/max': 1.4263579845428467, 'sampling/importance_sampling_ratio/min': 0.2401820868253708, 'sampling/importance_sampling_ratio/mean': 0.9938217997550964, 'sampling/importance_sampling_ratio/max': 1.8640339374542236, 'entropy': 0.4641291946172714, 'clip_ratio/low_mean': 0.019736841320991516, 'clip_ratio/low_min': 0.019736841320991516, 'clip_ratio/high_mean': 0.07001307792961597, 'clip_ratio/high_max': 0.07001307792961597, 'clip_ratio/region_mean': 0.08974991925060749, 'reward_total_mean': 0.24640285968780518, 'reward_meter_mean': 0.5552480220794678, 'reward_meter_std': 0.35175880789756775, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9113634824752808, 'reward_lexical_plausibility_std': 0.16448672115802765, 'reward_repeat_penalty_mean': 0.7949947118759155, 'reward_repeat_penalty_std': 0.21434663236141205, 'reward_near_duplicate_penalty_mean': 0.9128570556640625, 'reward_near_duplicate_penalty_std': 0.06426502019166946, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.8788214921951294, 'reward_distinct_2_std': 0.14814431965351105, 'reward_judge_quality_mean': 0.5375000238418579, 'reward_judge_quality_std': 0.3681517541408539, 'reward_total_composite_mean': 0.24640285968780518, 'reward_total_composite_std': 0.20807690918445587, 'epoch': 0.04}

 27%|██▋       | 900/3300 [3:36:00<8:51:38, 13.29s/it]INFO 04-12 14:22:12 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


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

 15%|█▌        | 2/13 [00:11<01:04,  5.87s/it]

 23%|██▎       | 3/13 [00:18<01:01,  6.18s/it]

 31%|███       | 4/13 [00:30<01:16,  8.45s/it]

 38%|███▊      | 5/13 [00:41<01:14,  9.29s/it]

 46%|████▌     | 6/13 [00:52<01:09,  9.87s/it]

 54%|█████▍    | 7/13 [01:04<01:04, 10.69s/it]

 62%|██████▏   | 8/13 [01:16<00:54, 10.98s/it]

 69%|██████▉   | 9/13 [01:27<00:44, 11.04s/it]

 77%|███████▋  | 10/13 [01:37<00:32, 10.78s/it]

 85%|████████▍ | 11/13 [01:48<00:21, 10.58s/it]

 92%|█████████▏| 12/13 [01:59<00:10, 10.73s/it]

100%|██████████| 13/13 [02:10<00:00, 10.86s/it]2026-04-12 14:24:35,052 | INFO | train_grpo_train | metrics_logged mode=eval step=900

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 142.7142, 'eval_samples_per_second': 0.729, 'eval_steps_per_second': 0.091, 'eval_num_tokens': 1832644.0, 'eval_completions/mean_length': 238.27884615384616, 'eval_completions/min_length': 48.07692307692308, 'eval_completions/max_length': 474.38461538461536, 'eval_completions/clipped_ratio': 0.18269230769230768, 'eval_completions/mean_terminated_length': 173.86658184344952, 'eval_completions/min_terminated_length': 48.07692307692308, 'eval_completions/max_terminated_length': 340.2307692307692, 'eval_rewards/meter/mean': 0.4154845716861578, 'eval_rewards/meter/std': 0.3006790383503987, 'eval_rewards/count_adherence/mean': 0.8665537696618301, 'eval_rewards/count_adherence/std': 0.18183370335743979, 'eval_rewards/arabic_clean/mean': 0.9134615384615384, 'eval_rewards/arabic_clean/std': 0.18420560084856474, 'eval_rewards/lexical_plausibility/mean': 0.9800223708152771, 'eval_rewards/lexical_plausibility/std': 0.04943215696570965, 'eval_rewards/judge_quality/mean': 0.28105768675987536, 'eval_rewards/judge_quality/std': 0.1960173398256302, 'eval_rewards/repeat_penalty/mean': 0.4591149366818942, 'eval_rewards/repeat_penalty/std': 0.3970482899592473, 'eval_rewards/near_duplicate_penalty/mean': 0.8061947134824899, 'eval_rewards/near_duplicate_penalty/std': 0.18771734489844397, 'eval_rewards/opening_diversity/mean': 0.9105327587861282, 'eval_rewards/opening_diversity/std': 0.17902263483175865, 'eval_rewards/distinct_2/mean': 0.5257952282061944, 'eval_rewards/distinct_2/std': 0.41410436767798203, 'eval_rewards/total_composite/mean': 0.11126948721133746, 'eval_rewards/total_composite/std': 0.13009287187686333, 'eval_reward': 0.11126948721133746, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.040275223553180695, 'eval_sampling/sampling_logp_difference/max': 1.0106278382814848, 'eval_sampling/importance_sampling_ratio/min': 0.3774389555821052, 'eval_sampling/importance_sampling_ratio/mean': 1.007907087986286, 'eval_sampling/importance_sampling_ratio/max': 1.4787569687916682, 'eval_entropy': 0.41509873362687916, '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': 0.11126948721133746, 'eval_reward_meter_mean': 0.4154845716861578, 'eval_reward_meter_std': 0.3006790383503987, 'eval_reward_count_adherence_mean': 0.8665537696618301, 'eval_reward_count_adherence_std': 0.18183370335743979, 'eval_reward_arabic_clean_mean': 0.9134615384615384, 'eval_reward_arabic_clean_std': 0.18420560084856474, 'eval_reward_lexical_plausibility_mean': 0.9800223708152771, 'eval_reward_lexical_plausibility_std': 0.04943215696570965, 'eval_reward_repeat_penalty_mean': 0.4591149366818942, 'eval_reward_repeat_penalty_std': 0.3970482899592473, 'eval_reward_near_duplicate_penalty_mean': 0.8061947134824899, 'eval_reward_near_duplicate_penalty_std': 0.18771734489844397, 'eval_reward_opening_diversity_mean': 0.9105327587861282, 'eval_reward_opening_diversity_std': 0.17902263483175865, 'eval_reward_distinct_2_mean': 0.5257952282061944, 'eval_reward_distinct_2_std': 0.41410436767798203, 'eval_reward_judge_quality_mean': 0.28105768675987536, 'eval_reward_judge_quality_std': 0.1960173398256302, 'eval_reward_total_composite_mean': 0.11126948721133746, 'eval_reward_total_composite_std': 0.13009287187686333, 'epoch': 0.04}

 27%|██▋       | 900/3300 [3:38:23<8:51:38, 13.29s/it]

100%|██████████| 13/13 [02:10<00:00, 10.86s/it]

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

 27%|██▋       | 901/3300 [3:38:39<37:58:54, 57.00s/it]2026-04-12 14:24:51,247 | INFO | train_grpo_train | metrics_logged mode=train step=901

                                                       
{'loss': -0.0466, 'grad_norm': 4.856034278869629, 'learning_rate': 7.272727272727273e-06, 'num_tokens': 1834218.0, 'completions/mean_length': 106.75, 'completions/min_length': 47.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 48.85714340209961, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.5247308015823364, 'rewards/meter/std': 0.41064566373825073, '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/lexical_plausibility/mean': 0.9642213582992554, 'rewards/lexical_plausibility/std': 0.10119719803333282, 'rewards/judge_quality/mean': 0.5262500047683716, 'rewards/judge_quality/std': 0.3414648175239563, 'rewards/repeat_penalty/mean': 0.9651439189910889, 'rewards/repeat_penalty/std': 0.059345826506614685, 'rewards/near_duplicate_penalty/mean': 0.9673627614974976, 'rewards/near_duplicate_penalty/std': 0.05336698517203331, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9973656535148621, 'rewards/distinct_2/std': 0.007451077923178673, 'rewards/total_composite/mean': 0.38585364818573, 'rewards/total_composite/std': 0.4109489619731903, 'reward': 0.38585364818573, 'reward_std': 0.4109489619731903, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10950043797492981, 'sampling/sampling_logp_difference/max': 1.1998834609985352, 'sampling/importance_sampling_ratio/min': 0.3012292981147766, 'sampling/importance_sampling_ratio/mean': 1.0183032751083374, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.48758653923869133, 'clip_ratio/low_mean': 0.057705594692379236, 'clip_ratio/low_min': 0.057705594692379236, 'clip_ratio/high_mean': 0.012655061902478337, 'clip_ratio/high_max': 0.012655061902478337, 'clip_ratio/region_mean': 0.07036065659485757, 'reward_total_mean': 0.38585364818573, 'reward_meter_mean': 0.5247308015823364, 'reward_meter_std': 0.41064566373825073, '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_lexical_plausibility_mean': 0.9642213582992554, 'reward_lexical_plausibility_std': 0.10119719803333282, 'reward_repeat_penalty_mean': 0.9651439189910889, 'reward_repeat_penalty_std': 0.059345826506614685, 'reward_near_duplicate_penalty_mean': 0.9673627614974976, 'reward_near_duplicate_penalty_std': 0.05336698517203331, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9973656535148621, 'reward_distinct_2_std': 0.007451077923178673, 'reward_judge_quality_mean': 0.5262500047683716, 'reward_judge_quality_std': 0.3414648175239563, 'reward_total_composite_mean': 0.38585364818573, 'reward_total_composite_std': 0.4109489619731903, 'epoch': 0.04}

 27%|██▋       | 901/3300 [3:38:39<37:58:54, 57.00s/it]INFO 04-12 14:24:51 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 902/3300 [3:38:52<29:12:38, 43.85s/it]2026-04-12 14:25:04,430 | INFO | train_grpo_train | metrics_logged mode=train step=902

                                                       
{'loss': -0.0786, 'grad_norm': 5.1406779289245605, 'learning_rate': 7.26969696969697e-06, 'num_tokens': 1836189.0, 'completions/mean_length': 134.375, 'completions/min_length': 67.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 80.42857360839844, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 91.0, 'rewards/meter/mean': 0.6462476849555969, 'rewards/meter/std': 0.3270873427391052, '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/lexical_plausibility/mean': 0.9690636992454529, 'rewards/lexical_plausibility/std': 0.08750104904174805, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.865140438079834, 'rewards/repeat_penalty/std': 0.11856569349765778, 'rewards/near_duplicate_penalty/mean': 0.9454061985015869, 'rewards/near_duplicate_penalty/std': 0.053187109529972076, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9117323160171509, 'rewards/distinct_2/std': 0.08192819356918335, 'rewards/total_composite/mean': 0.17186760902404785, 'rewards/total_composite/std': 0.11935804784297943, 'reward': 0.17186760902404785, 'reward_std': 0.11935804039239883, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13567602634429932, 'sampling/sampling_logp_difference/max': 1.5272159576416016, 'sampling/importance_sampling_ratio/min': 0.2171393483877182, 'sampling/importance_sampling_ratio/mean': 1.0106441974639893, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7391554228961468, 'clip_ratio/low_mean': 0.0436558430083096, 'clip_ratio/low_min': 0.0436558430083096, 'clip_ratio/high_mean': 0.0703059108927846, 'clip_ratio/high_max': 0.0703059108927846, 'clip_ratio/region_mean': 0.1139617539010942, 'reward_total_mean': 0.17186760902404785, 'reward_meter_mean': 0.6462476849555969, 'reward_meter_std': 0.3270873427391052, '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_lexical_plausibility_mean': 0.9690636992454529, 'reward_lexical_plausibility_std': 0.08750104904174805, 'reward_repeat_penalty_mean': 0.865140438079834, 'reward_repeat_penalty_std': 0.11856569349765778, 'reward_near_duplicate_penalty_mean': 0.9454061985015869, 'reward_near_duplicate_penalty_std': 0.053187109529972076, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9117323160171509, 'reward_distinct_2_std': 0.08192819356918335, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.17186760902404785, 'reward_total_composite_std': 0.11935804784297943, 'epoch': 0.04}

 27%|██▋       | 902/3300 [3:38:52<29:12:38, 43.85s/it]INFO 04-12 14:25:04 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 903/3300 [3:39:00<22:07:00, 33.22s/it]2026-04-12 14:25:12,825 | INFO | train_grpo_train | metrics_logged mode=train step=903

                                                       
{'loss': 0.0216, 'grad_norm': 7.0134782791137695, 'learning_rate': 7.266666666666668e-06, 'num_tokens': 1838181.0, 'completions/mean_length': 80.0, 'completions/min_length': 71.0, 'completions/max_length': 86.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.0, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.6698614358901978, 'rewards/meter/std': 0.3118594288825989, '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/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.0353553369641304, 'rewards/repeat_penalty/mean': 0.6805667877197266, 'rewards/repeat_penalty/std': 0.17560771107673645, 'rewards/near_duplicate_penalty/mean': 0.9137561917304993, 'rewards/near_duplicate_penalty/std': 0.03349655121564865, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8141463994979858, 'rewards/distinct_2/std': 0.10309074074029922, 'rewards/total_composite/mean': 0.22637386620044708, 'rewards/total_composite/std': 0.10816343128681183, 'reward': 0.22637386620044708, 'reward_std': 0.10816343128681183, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07582775503396988, 'sampling/sampling_logp_difference/max': 2.568497657775879, 'sampling/importance_sampling_ratio/min': 0.07665061205625534, 'sampling/importance_sampling_ratio/mean': 1.0025361776351929, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3810967281460762, 'clip_ratio/low_mean': 0.038531263009645045, 'clip_ratio/low_min': 0.038531263009645045, 'clip_ratio/high_mean': 0.03425446106120944, 'clip_ratio/high_max': 0.03425446106120944, 'clip_ratio/region_mean': 0.07278572407085449, 'reward_total_mean': 0.22637386620044708, 'reward_meter_mean': 0.6698614358901978, 'reward_meter_std': 0.3118594288825989, '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_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.6805667877197266, 'reward_repeat_penalty_std': 0.17560771107673645, 'reward_near_duplicate_penalty_mean': 0.9137561917304993, 'reward_near_duplicate_penalty_std': 0.03349655121564865, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8141463994979858, 'reward_distinct_2_std': 0.10309074074029922, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.0353553369641304, 'reward_total_composite_mean': 0.22637386620044708, 'reward_total_composite_std': 0.10816343128681183, 'epoch': 0.04}

 27%|██▋       | 903/3300 [3:39:00<22:07:00, 33.22s/it]INFO 04-12 14:25:13 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 904/3300 [3:39:14<18:12:22, 27.35s/it]2026-04-12 14:25:26,507 | INFO | train_grpo_train | metrics_logged mode=train step=904

                                                       
{'loss': -0.0198, 'grad_norm': 2.0801613330841064, 'learning_rate': 7.263636363636364e-06, 'num_tokens': 1839657.0, 'completions/mean_length': 202.5, 'completions/min_length': 15.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 16.80000114440918, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 19.0, 'rewards/meter/mean': 0.3084750771522522, 'rewards/meter/std': 0.4223470687866211, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.9139069318771362, 'rewards/lexical_plausibility/std': 0.11909569054841995, 'rewards/judge_quality/mean': 0.3050000071525574, 'rewards/judge_quality/std': 0.38239845633506775, 'rewards/repeat_penalty/mean': 0.7804979085922241, 'rewards/repeat_penalty/std': 0.08715314418077469, 'rewards/near_duplicate_penalty/mean': 0.9996334314346313, 'rewards/near_duplicate_penalty/std': 0.0010368125513195992, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9995228052139282, 'rewards/distinct_2/std': 0.0013496896717697382, 'rewards/total_composite/mean': 0.05777446925640106, 'rewards/total_composite/std': 0.07636607438325882, 'reward': 0.05777446925640106, 'reward_std': 0.07636607438325882, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2103031724691391, 'sampling/sampling_logp_difference/max': 2.262977123260498, 'sampling/importance_sampling_ratio/min': 0.10404027998447418, 'sampling/importance_sampling_ratio/mean': 0.9945732355117798, 'sampling/importance_sampling_ratio/max': 1.6447299718856812, 'entropy': 0.43526381626725197, 'clip_ratio/low_mean': 0.030330882407724857, 'clip_ratio/low_min': 0.030330882407724857, 'clip_ratio/high_mean': 0.08346233330667019, 'clip_ratio/high_max': 0.08346233330667019, 'clip_ratio/region_mean': 0.11379321571439505, 'reward_total_mean': 0.05777446925640106, 'reward_meter_mean': 0.3084750771522522, 'reward_meter_std': 0.4223470687866211, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.9139069318771362, 'reward_lexical_plausibility_std': 0.11909569054841995, 'reward_repeat_penalty_mean': 0.7804979085922241, 'reward_repeat_penalty_std': 0.08715314418077469, 'reward_near_duplicate_penalty_mean': 0.9996334314346313, 'reward_near_duplicate_penalty_std': 0.0010368125513195992, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9995228052139282, 'reward_distinct_2_std': 0.0013496896717697382, 'reward_judge_quality_mean': 0.3050000071525574, 'reward_judge_quality_std': 0.38239845633506775, 'reward_total_composite_mean': 0.05777446925640106, 'reward_total_composite_std': 0.07636607438325882, 'epoch': 0.04}

 27%|██▋       | 904/3300 [3:39:14<18:12:22, 27.35s/it]INFO 04-12 14:25:26 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 905/3300 [3:39:23<14:32:58, 21.87s/it]2026-04-12 14:25:35,577 | INFO | train_grpo_train | metrics_logged mode=train step=905

                                                       
{'loss': 0.0094, 'grad_norm': 10.159344673156738, 'learning_rate': 7.260606060606061e-06, 'num_tokens': 1841789.0, 'completions/mean_length': 103.5, 'completions/min_length': 97.0, 'completions/max_length': 111.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 103.5, 'completions/min_terminated_length': 97.0, 'completions/max_terminated_length': 111.0, 'rewards/meter/mean': 0.7616041898727417, 'rewards/meter/std': 0.27403607964515686, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/repeat_penalty/mean': 0.4398246705532074, 'rewards/repeat_penalty/std': 0.2828714847564697, 'rewards/near_duplicate_penalty/mean': 0.8676952123641968, 'rewards/near_duplicate_penalty/std': 0.07762881368398666, 'rewards/opening_diversity/mean': 0.5859375, 'rewards/opening_diversity/std': 0.2911214232444763, 'rewards/distinct_2/mean': 0.7911661863327026, 'rewards/distinct_2/std': 0.09991541504859924, 'rewards/total_composite/mean': 0.24073660373687744, 'rewards/total_composite/std': 0.11653447151184082, 'reward': 0.24073660373687744, 'reward_std': 0.11653447151184082, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10521750152111053, 'sampling/sampling_logp_difference/max': 2.861323118209839, 'sampling/importance_sampling_ratio/min': 0.05719303712248802, 'sampling/importance_sampling_ratio/mean': 0.9967947602272034, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4639259986579418, 'clip_ratio/low_mean': 0.024730920791625977, 'clip_ratio/low_min': 0.024730920791625977, 'clip_ratio/high_mean': 0.05875509697943926, 'clip_ratio/high_max': 0.05875509697943926, 'clip_ratio/region_mean': 0.08348601777106524, 'reward_total_mean': 0.24073660373687744, 'reward_meter_mean': 0.7616041898727417, 'reward_meter_std': 0.27403607964515686, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.4398246705532074, 'reward_repeat_penalty_std': 0.2828714847564697, 'reward_near_duplicate_penalty_mean': 0.8676952123641968, 'reward_near_duplicate_penalty_std': 0.07762881368398666, 'reward_opening_diversity_mean': 0.5859375, 'reward_opening_diversity_std': 0.2911214232444763, 'reward_distinct_2_mean': 0.7911661863327026, 'reward_distinct_2_std': 0.09991541504859924, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.24073660373687744, 'reward_total_composite_std': 0.11653447151184082, 'epoch': 0.04}

 27%|██▋       | 905/3300 [3:39:23<14:32:58, 21.87s/it]INFO 04-12 14:25:35 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 906/3300 [3:39:31<11:45:10, 17.67s/it]2026-04-12 14:25:43,460 | INFO | train_grpo_train | metrics_logged mode=train step=906

                                                       
{'loss': 0.0648, 'grad_norm': 17.03221321105957, 'learning_rate': 7.257575757575758e-06, 'num_tokens': 1843224.0, 'completions/mean_length': 24.375, 'completions/min_length': 21.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.375, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.6680982112884521, 'rewards/meter/std': 0.3179086148738861, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4962500035762787, 'rewards/judge_quality/std': 0.3767886161804199, 'rewards/repeat_penalty/mean': 0.7389006614685059, 'rewards/repeat_penalty/std': 0.024166369810700417, 'rewards/near_duplicate_penalty/mean': 0.9852008819580078, 'rewards/near_duplicate_penalty/std': 0.03222181275486946, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.27304381132125854, 'rewards/total_composite/std': 0.18368299305438995, 'reward': 0.27304381132125854, 'reward_std': 0.18368297815322876, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11624179035425186, 'sampling/sampling_logp_difference/max': 2.099454402923584, 'sampling/importance_sampling_ratio/min': 0.12252325564622879, 'sampling/importance_sampling_ratio/mean': 1.004376769065857, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5289847925305367, 'clip_ratio/low_mean': 0.038920940831303596, 'clip_ratio/low_min': 0.038920940831303596, 'clip_ratio/high_mean': 0.03754787798970938, 'clip_ratio/high_max': 0.03754787798970938, 'clip_ratio/region_mean': 0.07646881882101297, 'reward_total_mean': 0.27304381132125854, 'reward_meter_mean': 0.6680982112884521, 'reward_meter_std': 0.3179086148738861, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7389006614685059, 'reward_repeat_penalty_std': 0.024166369810700417, 'reward_near_duplicate_penalty_mean': 0.9852008819580078, 'reward_near_duplicate_penalty_std': 0.03222181275486946, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4962500035762787, 'reward_judge_quality_std': 0.3767886161804199, 'reward_total_composite_mean': 0.27304381132125854, 'reward_total_composite_std': 0.18368299305438995, 'epoch': 0.04}

 27%|██▋       | 906/3300 [3:39:31<11:45:10, 17.67s/it]INFO 04-12 14:25:43 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 907/3300 [3:39:42<10:22:04, 15.60s/it]2026-04-12 14:25:54,214 | INFO | train_grpo_train | metrics_logged mode=train step=907

                                                       
{'loss': 0.0516, 'grad_norm': 7.356684684753418, 'learning_rate': 7.254545454545455e-06, 'num_tokens': 1846165.0, 'completions/mean_length': 177.625, 'completions/min_length': 158.0, 'completions/max_length': 215.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 177.625, 'completions/min_terminated_length': 158.0, 'completions/max_terminated_length': 215.0, 'rewards/meter/mean': 0.26473280787467957, 'rewards/meter/std': 0.2548903524875641, 'rewards/count_adherence/mean': 0.9444444179534912, 'rewards/count_adherence/std': 0.059391383081674576, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.10690448433160782, 'rewards/repeat_penalty/mean': 0.739059567451477, 'rewards/repeat_penalty/std': 0.21363458037376404, 'rewards/near_duplicate_penalty/mean': 0.9342135787010193, 'rewards/near_duplicate_penalty/std': 0.03609912469983101, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8026244640350342, 'rewards/distinct_2/std': 0.183920755982399, 'rewards/total_composite/mean': 0.06867167353630066, 'rewards/total_composite/std': 0.0791734829545021, 'reward': 0.06867167353630066, 'reward_std': 0.0791734829545021, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11483689397573471, 'sampling/sampling_logp_difference/max': 3.0560803413391113, 'sampling/importance_sampling_ratio/min': 0.04707184061408043, 'sampling/importance_sampling_ratio/mean': 1.0007380247116089, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5057728476822376, 'clip_ratio/low_mean': 0.048943258821964264, 'clip_ratio/low_min': 0.048943258821964264, 'clip_ratio/high_mean': 0.06487262714654207, 'clip_ratio/high_max': 0.06487262714654207, 'clip_ratio/region_mean': 0.11381588596850634, 'reward_total_mean': 0.06867167353630066, 'reward_meter_mean': 0.26473280787467957, 'reward_meter_std': 0.2548903524875641, 'reward_count_adherence_mean': 0.9444444179534912, 'reward_count_adherence_std': 0.059391383081674576, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.739059567451477, 'reward_repeat_penalty_std': 0.21363458037376404, 'reward_near_duplicate_penalty_mean': 0.9342135787010193, 'reward_near_duplicate_penalty_std': 0.03609912469983101, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8026244640350342, 'reward_distinct_2_std': 0.183920755982399, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.10690448433160782, 'reward_total_composite_mean': 0.06867167353630066, 'reward_total_composite_std': 0.0791734829545021, 'epoch': 0.04}

 27%|██▋       | 907/3300 [3:39:42<10:22:04, 15.60s/it]INFO 04-12 14:25:54 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 908/3300 [3:39:55<9:59:22, 15.03s/it] 2026-04-12 14:26:07,933 | INFO | train_grpo_train | metrics_logged mode=train step=908

                                                      
{'loss': -0.0504, 'grad_norm': 2.2303521633148193, 'learning_rate': 7.251515151515151e-06, 'num_tokens': 1847669.0, 'completions/mean_length': 220.0, 'completions/min_length': 41.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 44.79999923706055, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.46474987268447876, 'rewards/meter/std': 0.4497232139110565, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.8987423181533813, 'rewards/lexical_plausibility/std': 0.13027387857437134, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.19955307245254517, 'rewards/repeat_penalty/mean': 0.8320578336715698, 'rewards/repeat_penalty/std': 0.13103018701076508, 'rewards/near_duplicate_penalty/mean': 0.9724264740943909, 'rewards/near_duplicate_penalty/std': 0.025230664759874344, 'rewards/opening_diversity/mean': 0.875, 'rewards/opening_diversity/std': 0.13363061845302582, 'rewards/distinct_2/mean': 0.9782222509384155, 'rewards/distinct_2/std': 0.03798056021332741, 'rewards/total_composite/mean': 0.1411353200674057, 'rewards/total_composite/std': 0.18911324441432953, 'reward': 0.1411353200674057, 'reward_std': 0.18911321461200714, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11839240789413452, 'sampling/sampling_logp_difference/max': 1.9321202039718628, 'sampling/importance_sampling_ratio/min': 0.1448407769203186, 'sampling/importance_sampling_ratio/mean': 0.999984085559845, 'sampling/importance_sampling_ratio/max': 1.786693811416626, 'entropy': 0.3215724155306816, 'clip_ratio/low_mean': 0.03264146577566862, 'clip_ratio/low_min': 0.03264146577566862, 'clip_ratio/high_mean': 0.03963867202401161, 'clip_ratio/high_max': 0.03963867202401161, 'clip_ratio/region_mean': 0.07228013779968023, 'reward_total_mean': 0.1411353200674057, 'reward_meter_mean': 0.46474987268447876, 'reward_meter_std': 0.4497232139110565, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.8987423181533813, 'reward_lexical_plausibility_std': 0.13027387857437134, 'reward_repeat_penalty_mean': 0.8320578336715698, 'reward_repeat_penalty_std': 0.13103018701076508, 'reward_near_duplicate_penalty_mean': 0.9724264740943909, 'reward_near_duplicate_penalty_std': 0.025230664759874344, 'reward_opening_diversity_mean': 0.875, 'reward_opening_diversity_std': 0.13363061845302582, 'reward_distinct_2_mean': 0.9782222509384155, 'reward_distinct_2_std': 0.03798056021332741, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.19955307245254517, 'reward_total_composite_mean': 0.1411353200674057, 'reward_total_composite_std': 0.18911324441432953, 'epoch': 0.04}

 28%|██▊       | 908/3300 [3:39:55<9:59:22, 15.03s/it]INFO 04-12 14:26:08 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 909/3300 [3:40:04<8:45:48, 13.19s/it]2026-04-12 14:26:16,836 | INFO | train_grpo_train | metrics_logged mode=train step=909

                                                      
{'loss': 0.0006, 'grad_norm': 10.0469331741333, 'learning_rate': 7.2484848484848495e-06, 'num_tokens': 1849818.0, 'completions/mean_length': 90.625, 'completions/min_length': 83.0, 'completions/max_length': 100.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 90.625, 'completions/min_terminated_length': 83.0, 'completions/max_terminated_length': 100.0, 'rewards/meter/mean': 0.46023374795913696, 'rewards/meter/std': 0.33316048979759216, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.8125627636909485, 'rewards/repeat_penalty/std': 0.12995557487010956, 'rewards/near_duplicate_penalty/mean': 0.9248131513595581, 'rewards/near_duplicate_penalty/std': 0.04308905452489853, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8911466598510742, 'rewards/distinct_2/std': 0.09683087468147278, 'rewards/total_composite/mean': 0.1353432834148407, 'rewards/total_composite/std': 0.11111269891262054, 'reward': 0.1353432834148407, 'reward_std': 0.11111269891262054, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11659479141235352, 'sampling/sampling_logp_difference/max': 3.9343385696411133, 'sampling/importance_sampling_ratio/min': 0.019558632746338844, 'sampling/importance_sampling_ratio/mean': 0.9950460195541382, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4510437734425068, 'clip_ratio/low_mean': 0.04797249333932996, 'clip_ratio/low_min': 0.04797249333932996, 'clip_ratio/high_mean': 0.042016807943582535, 'clip_ratio/high_max': 0.042016807943582535, 'clip_ratio/region_mean': 0.08998930128291249, 'reward_total_mean': 0.1353432834148407, 'reward_meter_mean': 0.46023374795913696, 'reward_meter_std': 0.33316048979759216, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8125627636909485, 'reward_repeat_penalty_std': 0.12995557487010956, 'reward_near_duplicate_penalty_mean': 0.9248131513595581, 'reward_near_duplicate_penalty_std': 0.04308905452489853, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8911466598510742, 'reward_distinct_2_std': 0.09683087468147278, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.1353432834148407, 'reward_total_composite_std': 0.11111269891262054, 'epoch': 0.04}

 28%|██▊       | 909/3300 [3:40:04<8:45:48, 13.19s/it]INFO 04-12 14:26:17 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 910/3300 [3:40:18<8:55:54, 13.45s/it]2026-04-12 14:26:30,893 | INFO | train_grpo_train | metrics_logged mode=train step=910

                                                      
{'loss': -0.0407, 'grad_norm': 3.7108423709869385, 'learning_rate': 7.245454545454546e-06, 'num_tokens': 1853614.0, 'completions/mean_length': 297.5, 'completions/min_length': 208.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 266.8571472167969, 'completions/min_terminated_length': 208.0, 'completions/max_terminated_length': 341.0, 'rewards/meter/mean': 0.26572930812835693, 'rewards/meter/std': 0.20120589435100555, 'rewards/count_adherence/mean': 0.71875, 'rewards/count_adherence/std': 0.2086307406425476, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.961482584476471, 'rewards/lexical_plausibility/std': 0.10894373059272766, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.12464234977960587, 'rewards/repeat_penalty/mean': 0.4588620960712433, 'rewards/repeat_penalty/std': 0.2753836512565613, 'rewards/near_duplicate_penalty/mean': 0.876366376876831, 'rewards/near_duplicate_penalty/std': 0.0673738643527031, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.5448314547538757, 'rewards/distinct_2/std': 0.2569446265697479, 'rewards/total_composite/mean': 0.04977791756391525, 'rewards/total_composite/std': 0.05664224550127983, 'reward': 0.04977791756391525, 'reward_std': 0.05664224550127983, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09329218417406082, 'sampling/sampling_logp_difference/max': 3.3235740661621094, 'sampling/importance_sampling_ratio/min': 0.03602385148406029, 'sampling/importance_sampling_ratio/mean': 1.000282645225525, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3965323455631733, 'clip_ratio/low_mean': 0.029428347013890743, 'clip_ratio/low_min': 0.029428347013890743, 'clip_ratio/high_mean': 0.056563181802630424, 'clip_ratio/high_max': 0.056563181802630424, 'clip_ratio/region_mean': 0.08599152881652117, 'reward_total_mean': 0.04977791756391525, 'reward_meter_mean': 0.26572930812835693, 'reward_meter_std': 0.20120589435100555, 'reward_count_adherence_mean': 0.71875, 'reward_count_adherence_std': 0.2086307406425476, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.961482584476471, 'reward_lexical_plausibility_std': 0.10894373059272766, 'reward_repeat_penalty_mean': 0.4588620960712433, 'reward_repeat_penalty_std': 0.2753836512565613, 'reward_near_duplicate_penalty_mean': 0.876366376876831, 'reward_near_duplicate_penalty_std': 0.0673738643527031, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.5448314547538757, 'reward_distinct_2_std': 0.2569446265697479, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.12464234977960587, 'reward_total_composite_mean': 0.04977791756391525, 'reward_total_composite_std': 0.05664224550127983, 'epoch': 0.04}

 28%|██▊       | 910/3300 [3:40:18<8:55:54, 13.45s/it]INFO 04-12 14:26:31 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 911/3300 [3:40:27<7:56:57, 11.98s/it]2026-04-12 14:26:39,430 | INFO | train_grpo_train | metrics_logged mode=train step=911

                                                      
{'loss': -0.0177, 'grad_norm': 13.395476341247559, 'learning_rate': 7.242424242424243e-06, 'num_tokens': 1855236.0, 'completions/mean_length': 55.75, 'completions/min_length': 45.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.75, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.3058864176273346, 'rewards/meter/std': 0.3473019599914551, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5550000071525574, 'rewards/judge_quality/std': 0.22790977358818054, 'rewards/repeat_penalty/mean': 0.9301947355270386, 'rewards/repeat_penalty/std': 0.047629062086343765, 'rewards/near_duplicate_penalty/mean': 0.9638501405715942, 'rewards/near_duplicate_penalty/std': 0.027497611939907074, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9649626016616821, 'rewards/distinct_2/std': 0.037533897906541824, 'rewards/total_composite/mean': 0.17312057316303253, 'rewards/total_composite/std': 0.1900882124900818, 'reward': 0.17312057316303253, 'reward_std': 0.1900882124900818, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1404111683368683, 'sampling/sampling_logp_difference/max': 1.8854422569274902, 'sampling/importance_sampling_ratio/min': 0.15176191926002502, 'sampling/importance_sampling_ratio/mean': 1.002879023551941, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7668841406702995, 'clip_ratio/low_mean': 0.08556771744042635, 'clip_ratio/low_min': 0.08556771744042635, 'clip_ratio/high_mean': 0.0555866751819849, 'clip_ratio/high_max': 0.0555866751819849, 'clip_ratio/region_mean': 0.14115439262241125, 'reward_total_mean': 0.17312057316303253, 'reward_meter_mean': 0.3058864176273346, 'reward_meter_std': 0.3473019599914551, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9301947355270386, 'reward_repeat_penalty_std': 0.047629062086343765, 'reward_near_duplicate_penalty_mean': 0.9638501405715942, 'reward_near_duplicate_penalty_std': 0.027497611939907074, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9649626016616821, 'reward_distinct_2_std': 0.037533897906541824, 'reward_judge_quality_mean': 0.5550000071525574, 'reward_judge_quality_std': 0.22790977358818054, 'reward_total_composite_mean': 0.17312057316303253, 'reward_total_composite_std': 0.1900882124900818, 'epoch': 0.04}

 28%|██▊       | 911/3300 [3:40:27<7:56:57, 11.98s/it]INFO 04-12 14:26:39 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 912/3300 [3:40:40<8:10:28, 12.32s/it]2026-04-12 14:26:52,559 | INFO | train_grpo_train | metrics_logged mode=train step=912

                                                      
{'loss': -0.0719, 'grad_norm': 3.359111785888672, 'learning_rate': 7.2393939393939404e-06, 'num_tokens': 1857468.0, 'completions/mean_length': 218.0, 'completions/min_length': 109.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 120.0, 'completions/min_terminated_length': 109.0, 'completions/max_terminated_length': 132.0, 'rewards/meter/mean': 0.22825634479522705, 'rewards/meter/std': 0.3028012216091156, 'rewards/count_adherence/mean': 0.8500000238418579, 'rewards/count_adherence/std': 0.2976095378398895, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9280493855476379, 'rewards/lexical_plausibility/std': 0.13533242046833038, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.155264750123024, 'rewards/repeat_penalty/mean': 0.7395656108856201, 'rewards/repeat_penalty/std': 0.22323991358280182, 'rewards/near_duplicate_penalty/mean': 0.9454048871994019, 'rewards/near_duplicate_penalty/std': 0.060150161385536194, 'rewards/opening_diversity/mean': 0.875, 'rewards/opening_diversity/std': 0.155838742852211, 'rewards/distinct_2/mean': 0.8757848143577576, 'rewards/distinct_2/std': 0.13074222207069397, 'rewards/total_composite/mean': 0.07523701339960098, 'rewards/total_composite/std': 0.10943896323442459, 'reward': 0.07523701339960098, 'reward_std': 0.10943896323442459, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1248907521367073, 'sampling/sampling_logp_difference/max': 1.9808249473571777, 'sampling/importance_sampling_ratio/min': 0.13795538246631622, 'sampling/importance_sampling_ratio/mean': 0.9841961860656738, 'sampling/importance_sampling_ratio/max': 1.9988473653793335, 'entropy': 0.3965604715049267, 'clip_ratio/low_mean': 0.0460313456133008, 'clip_ratio/low_min': 0.0460313456133008, 'clip_ratio/high_mean': 0.03056826815009117, 'clip_ratio/high_max': 0.03056826815009117, 'clip_ratio/region_mean': 0.07659961376339197, 'reward_total_mean': 0.07523701339960098, 'reward_meter_mean': 0.22825634479522705, 'reward_meter_std': 0.3028012216091156, 'reward_count_adherence_mean': 0.8500000238418579, 'reward_count_adherence_std': 0.2976095378398895, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9280493855476379, 'reward_lexical_plausibility_std': 0.13533242046833038, 'reward_repeat_penalty_mean': 0.7395656108856201, 'reward_repeat_penalty_std': 0.22323991358280182, 'reward_near_duplicate_penalty_mean': 0.9454048871994019, 'reward_near_duplicate_penalty_std': 0.060150161385536194, 'reward_opening_diversity_mean': 0.875, 'reward_opening_diversity_std': 0.155838742852211, 'reward_distinct_2_mean': 0.8757848143577576, 'reward_distinct_2_std': 0.13074222207069397, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.155264750123024, 'reward_total_composite_mean': 0.07523701339960098, 'reward_total_composite_std': 0.10943896323442459, 'epoch': 0.04}

 28%|██▊       | 912/3300 [3:40:40<8:10:28, 12.32s/it]INFO 04-12 14:26:52 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 913/3300 [3:40:54<8:35:02, 12.95s/it]2026-04-12 14:27:06,959 | INFO | train_grpo_train | metrics_logged mode=train step=913

                                                      
{'loss': 0.1245, 'grad_norm': 1.2069638967514038, 'learning_rate': 7.236363636363637e-06, 'num_tokens': 1859761.0, 'completions/mean_length': 507.625, 'completions/min_length': 477.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.875, 'completions/mean_terminated_length': 477.0, 'completions/min_terminated_length': 477.0, 'completions/max_terminated_length': 477.0, 'rewards/meter/mean': 0.35045263171195984, 'rewards/meter/std': 0.2506626546382904, 'rewards/count_adherence/mean': 0.5340909361839294, 'rewards/count_adherence/std': 0.09009373933076859, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.0037082768976688385, 'rewards/repeat_penalty/std': 0.00762829277664423, 'rewards/near_duplicate_penalty/mean': 0.7244557738304138, 'rewards/near_duplicate_penalty/std': 0.13442681729793549, 'rewards/opening_diversity/mean': 0.9689275622367859, 'rewards/opening_diversity/std': 0.07131610810756683, 'rewards/distinct_2/mean': 0.005805698223412037, 'rewards/distinct_2/std': 0.012150234542787075, 'rewards/total_composite/mean': 0.02310771867632866, 'rewards/total_composite/std': 0.02202078141272068, 'reward': 0.02310771867632866, 'reward_std': 0.02202078141272068, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05831579491496086, 'sampling/sampling_logp_difference/max': 1.9486279487609863, 'sampling/importance_sampling_ratio/min': 0.1424694061279297, 'sampling/importance_sampling_ratio/mean': 1.0042625665664673, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.026542680338025093, 'clip_ratio/low_mean': 0.006813417188823223, 'clip_ratio/low_min': 0.006813417188823223, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.006813417188823223, 'reward_total_mean': 0.02310771867632866, 'reward_meter_mean': 0.35045263171195984, 'reward_meter_std': 0.2506626546382904, 'reward_count_adherence_mean': 0.5340909361839294, 'reward_count_adherence_std': 0.09009373933076859, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.0037082768976688385, 'reward_repeat_penalty_std': 0.00762829277664423, 'reward_near_duplicate_penalty_mean': 0.7244557738304138, 'reward_near_duplicate_penalty_std': 0.13442681729793549, 'reward_opening_diversity_mean': 0.9689275622367859, 'reward_opening_diversity_std': 0.07131610810756683, 'reward_distinct_2_mean': 0.005805698223412037, 'reward_distinct_2_std': 0.012150234542787075, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.02310771867632866, 'reward_total_composite_std': 0.02202078141272068, 'epoch': 0.04}

 28%|██▊       | 913/3300 [3:40:54<8:35:02, 12.95s/it]INFO 04-12 14:27:07 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 914/3300 [3:41:08<8:47:14, 13.26s/it]2026-04-12 14:27:20,946 | INFO | train_grpo_train | metrics_logged mode=train step=914

                                                      
{'loss': -0.0787, 'grad_norm': 3.3468692302703857, 'learning_rate': 7.233333333333334e-06, 'num_tokens': 1861285.0, 'completions/mean_length': 108.5, 'completions/min_length': 49.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 50.85714340209961, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.8519883155822754, 'rewards/meter/std': 0.16690686345100403, '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/lexical_plausibility/mean': 0.9645233750343323, 'rewards/lexical_plausibility/std': 0.1003430113196373, 'rewards/judge_quality/mean': 0.20000000298023224, 'rewards/judge_quality/std': 0.10690449178218842, 'rewards/repeat_penalty/mean': 0.5674265027046204, 'rewards/repeat_penalty/std': 0.2832561433315277, 'rewards/near_duplicate_penalty/mean': 0.7704793810844421, 'rewards/near_duplicate_penalty/std': 0.1786900758743286, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.7631553411483765, 'rewards/distinct_2/std': 0.18417690694332123, 'rewards/total_composite/mean': 0.1603967249393463, 'rewards/total_composite/std': 0.10251737385988235, 'reward': 0.1603967249393463, 'reward_std': 0.10251737385988235, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09863579273223877, 'sampling/sampling_logp_difference/max': 1.2791244983673096, 'sampling/importance_sampling_ratio/min': 0.2782808244228363, 'sampling/importance_sampling_ratio/mean': 1.0042524337768555, 'sampling/importance_sampling_ratio/max': 1.9618065357208252, 'entropy': 0.49080899357795715, 'clip_ratio/low_mean': 0.04609534703195095, 'clip_ratio/low_min': 0.04609534703195095, 'clip_ratio/high_mean': 0.05709380004554987, 'clip_ratio/high_max': 0.05709380004554987, 'clip_ratio/region_mean': 0.10318914707750082, 'reward_total_mean': 0.1603967249393463, 'reward_meter_mean': 0.8519883155822754, 'reward_meter_std': 0.16690686345100403, '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_lexical_plausibility_mean': 0.9645233750343323, 'reward_lexical_plausibility_std': 0.1003430113196373, 'reward_repeat_penalty_mean': 0.5674265027046204, 'reward_repeat_penalty_std': 0.2832561433315277, 'reward_near_duplicate_penalty_mean': 0.7704793810844421, 'reward_near_duplicate_penalty_std': 0.1786900758743286, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.7631553411483765, 'reward_distinct_2_std': 0.18417690694332123, 'reward_judge_quality_mean': 0.20000000298023224, 'reward_judge_quality_std': 0.10690449178218842, 'reward_total_composite_mean': 0.1603967249393463, 'reward_total_composite_std': 0.10251737385988235, 'epoch': 0.04}

 28%|██▊       | 914/3300 [3:41:08<8:47:14, 13.26s/it]INFO 04-12 14:27:21 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 915/3300 [3:41:22<8:49:28, 13.32s/it]2026-04-12 14:27:34,412 | INFO | train_grpo_train | metrics_logged mode=train step=915

                                                      
{'loss': 0.0211, 'grad_norm': 3.023385763168335, 'learning_rate': 7.2303030303030305e-06, 'num_tokens': 1865865.0, 'completions/mean_length': 357.5, 'completions/min_length': 305.0, 'completions/max_length': 494.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 357.5, 'completions/min_terminated_length': 305.0, 'completions/max_terminated_length': 494.0, 'rewards/meter/mean': 0.3904086649417877, 'rewards/meter/std': 0.23703508079051971, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.09848947077989578, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.16250000894069672, 'rewards/judge_quality/std': 0.0353553369641304, 'rewards/repeat_penalty/mean': 0.09513363242149353, 'rewards/repeat_penalty/std': 0.10595035552978516, 'rewards/near_duplicate_penalty/mean': 0.7053005695343018, 'rewards/near_duplicate_penalty/std': 0.13026690483093262, 'rewards/opening_diversity/mean': 0.8410183191299438, 'rewards/opening_diversity/std': 0.24256747961044312, 'rewards/distinct_2/mean': 0.13615761697292328, 'rewards/distinct_2/std': 0.15210334956645966, 'rewards/total_composite/mean': 0.04659539461135864, 'rewards/total_composite/std': 0.028762316331267357, 'reward': 0.04659539461135864, 'reward_std': 0.028762316331267357, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05945450812578201, 'sampling/sampling_logp_difference/max': 1.8852674961090088, 'sampling/importance_sampling_ratio/min': 0.15178844332695007, 'sampling/importance_sampling_ratio/mean': 1.0019183158874512, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.298257851973176, 'clip_ratio/low_mean': 0.022474889643490314, 'clip_ratio/low_min': 0.022474889643490314, 'clip_ratio/high_mean': 0.04105802858248353, 'clip_ratio/high_max': 0.04105802858248353, 'clip_ratio/region_mean': 0.06353291822597384, 'reward_total_mean': 0.04659539461135864, 'reward_meter_mean': 0.3904086649417877, 'reward_meter_std': 0.23703508079051971, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.09848947077989578, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.09513363242149353, 'reward_repeat_penalty_std': 0.10595035552978516, 'reward_near_duplicate_penalty_mean': 0.7053005695343018, 'reward_near_duplicate_penalty_std': 0.13026690483093262, 'reward_opening_diversity_mean': 0.8410183191299438, 'reward_opening_diversity_std': 0.24256747961044312, 'reward_distinct_2_mean': 0.13615761697292328, 'reward_distinct_2_std': 0.15210334956645966, 'reward_judge_quality_mean': 0.16250000894069672, 'reward_judge_quality_std': 0.0353553369641304, 'reward_total_composite_mean': 0.04659539461135864, 'reward_total_composite_std': 0.028762316331267357, 'epoch': 0.04}

 28%|██▊       | 915/3300 [3:41:22<8:49:28, 13.32s/it]INFO 04-12 14:27:34 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 916/3300 [3:41:30<7:44:35, 11.69s/it]2026-04-12 14:27:42,301 | INFO | train_grpo_train | metrics_logged mode=train step=916

                                                      
{'loss': 0.0113, 'grad_norm': 12.755487442016602, 'learning_rate': 7.227272727272729e-06, 'num_tokens': 1867292.0, 'completions/mean_length': 34.375, 'completions/min_length': 32.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.375, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.662441611289978, 'rewards/meter/std': 0.3464275896549225, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4624999761581421, 'rewards/judge_quality/std': 0.08345229178667068, 'rewards/repeat_penalty/mean': 0.9969915151596069, 'rewards/repeat_penalty/std': 0.0043798936530947685, 'rewards/near_duplicate_penalty/mean': 0.9969915151596069, 'rewards/near_duplicate_penalty/std': 0.0043798936530947685, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.32099583745002747, 'rewards/total_composite/std': 0.1858326494693756, 'reward': 0.32099583745002747, 'reward_std': 0.1858326494693756, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12480656057596207, 'sampling/sampling_logp_difference/max': 1.4419769048690796, 'sampling/importance_sampling_ratio/min': 0.2364598512649536, 'sampling/importance_sampling_ratio/mean': 1.016209363937378, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.568106759339571, 'clip_ratio/low_mean': 0.051674108020961285, 'clip_ratio/low_min': 0.051674108020961285, 'clip_ratio/high_mean': 0.06877483334392309, 'clip_ratio/high_max': 0.06877483334392309, 'clip_ratio/region_mean': 0.12044894136488438, 'reward_total_mean': 0.32099583745002747, 'reward_meter_mean': 0.662441611289978, 'reward_meter_std': 0.3464275896549225, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9969915151596069, 'reward_repeat_penalty_std': 0.0043798936530947685, 'reward_near_duplicate_penalty_mean': 0.9969915151596069, 'reward_near_duplicate_penalty_std': 0.0043798936530947685, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4624999761581421, 'reward_judge_quality_std': 0.08345229178667068, 'reward_total_composite_mean': 0.32099583745002747, 'reward_total_composite_std': 0.1858326494693756, 'epoch': 0.04}

 28%|██▊       | 916/3300 [3:41:30<7:44:35, 11.69s/it]INFO 04-12 14:27:42 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 917/3300 [3:41:38<7:01:02, 10.60s/it]2026-04-12 14:27:50,358 | INFO | train_grpo_train | metrics_logged mode=train step=917

                                                      
{'loss': 0.0279, 'grad_norm': 15.212775230407715, 'learning_rate': 7.224242424242425e-06, 'num_tokens': 1868871.0, 'completions/mean_length': 45.375, 'completions/min_length': 42.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.375, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.26263055205345154, 'rewards/meter/std': 0.2262229025363922, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.9237500429153442, 'rewards/judge_quality/std': 0.01060659158974886, 'rewards/repeat_penalty/mean': 0.9845319986343384, 'rewards/repeat_penalty/std': 0.009524662047624588, 'rewards/near_duplicate_penalty/mean': 0.9845319986343384, 'rewards/near_duplicate_penalty/std': 0.009524662047624588, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.241818368434906, 'rewards/total_composite/std': 0.20791561901569366, 'reward': 0.241818368434906, 'reward_std': 0.20791560411453247, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1435861587524414, 'sampling/sampling_logp_difference/max': 1.8963546752929688, 'sampling/importance_sampling_ratio/min': 0.15011483430862427, 'sampling/importance_sampling_ratio/mean': 0.9993740320205688, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.49711066111922264, 'clip_ratio/low_mean': 0.06246612127870321, 'clip_ratio/low_min': 0.06246612127870321, 'clip_ratio/high_mean': 0.06104932352900505, 'clip_ratio/high_max': 0.06104932352900505, 'clip_ratio/region_mean': 0.12351544480770826, 'reward_total_mean': 0.241818368434906, 'reward_meter_mean': 0.26263055205345154, 'reward_meter_std': 0.2262229025363922, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9845319986343384, 'reward_repeat_penalty_std': 0.009524662047624588, 'reward_near_duplicate_penalty_mean': 0.9845319986343384, 'reward_near_duplicate_penalty_std': 0.009524662047624588, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.9237500429153442, 'reward_judge_quality_std': 0.01060659158974886, 'reward_total_composite_mean': 0.241818368434906, 'reward_total_composite_std': 0.20791561901569366, 'epoch': 0.04}

 28%|██▊       | 917/3300 [3:41:38<7:01:02, 10.60s/it]INFO 04-12 14:27:50 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 918/3300 [3:41:46<6:28:36,  9.79s/it]2026-04-12 14:27:58,248 | INFO | train_grpo_train | metrics_logged mode=train step=918

                                                      
{'loss': 0.2797, 'grad_norm': 18.46232032775879, 'learning_rate': 7.221212121212122e-06, 'num_tokens': 1870302.0, 'completions/mean_length': 30.875, 'completions/min_length': 24.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.875, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.8715851902961731, 'rewards/meter/std': 0.2659103274345398, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7100000381469727, 'rewards/judge_quality/std': 0.3123185336589813, 'rewards/repeat_penalty/mean': 0.7571017146110535, 'rewards/repeat_penalty/std': 0.07671567052602768, 'rewards/near_duplicate_penalty/mean': 0.9705473780632019, 'rewards/near_duplicate_penalty/std': 0.039805445820093155, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6072956323623657, 'rewards/total_composite/std': 0.4061472713947296, 'reward': 0.6072956323623657, 'reward_std': 0.4061472713947296, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14646393060684204, 'sampling/sampling_logp_difference/max': 2.408942461013794, 'sampling/importance_sampling_ratio/min': 0.08991033583879471, 'sampling/importance_sampling_ratio/mean': 0.9955363273620605, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7872043140232563, 'clip_ratio/low_mean': 0.029315476771444082, 'clip_ratio/low_min': 0.029315476771444082, 'clip_ratio/high_mean': 0.06673535006120801, 'clip_ratio/high_max': 0.06673535006120801, 'clip_ratio/region_mean': 0.09605082683265209, 'reward_total_mean': 0.6072956323623657, 'reward_meter_mean': 0.8715851902961731, 'reward_meter_std': 0.2659103274345398, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7571017146110535, 'reward_repeat_penalty_std': 0.07671567052602768, 'reward_near_duplicate_penalty_mean': 0.9705473780632019, 'reward_near_duplicate_penalty_std': 0.039805445820093155, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7100000381469727, 'reward_judge_quality_std': 0.3123185336589813, 'reward_total_composite_mean': 0.6072956323623657, 'reward_total_composite_std': 0.4061472713947296, 'epoch': 0.04}

 28%|██▊       | 918/3300 [3:41:46<6:28:36,  9.79s/it]INFO 04-12 14:27:58 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 919/3300 [3:41:54<6:10:36,  9.34s/it]2026-04-12 14:28:06,538 | INFO | train_grpo_train | metrics_logged mode=train step=919

                                                      
{'loss': 0.037, 'grad_norm': 13.303698539733887, 'learning_rate': 7.218181818181819e-06, 'num_tokens': 1872083.0, 'completions/mean_length': 55.625, 'completions/min_length': 50.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.625, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.7657922506332397, 'rewards/meter/std': 0.3170587420463562, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5262500047683716, 'rewards/judge_quality/std': 0.3414648175239563, 'rewards/repeat_penalty/mean': 0.9681443572044373, 'rewards/repeat_penalty/std': 0.040724486112594604, 'rewards/near_duplicate_penalty/mean': 0.9778196811676025, 'rewards/near_duplicate_penalty/std': 0.020623713731765747, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9898785352706909, 'rewards/distinct_2/std': 0.028627805411815643, 'rewards/total_composite/mean': 0.4392695426940918, 'rewards/total_composite/std': 0.35829058289527893, 'reward': 0.4392695426940918, 'reward_std': 0.35829055309295654, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12620609998703003, 'sampling/sampling_logp_difference/max': 1.7419829368591309, 'sampling/importance_sampling_ratio/min': 0.17517270147800446, 'sampling/importance_sampling_ratio/mean': 0.9975568056106567, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.591563381254673, 'clip_ratio/low_mean': 0.07790672034025192, 'clip_ratio/low_min': 0.07790672034025192, 'clip_ratio/high_mean': 0.04094973113387823, 'clip_ratio/high_max': 0.04094973113387823, 'clip_ratio/region_mean': 0.11885645147413015, 'reward_total_mean': 0.4392695426940918, 'reward_meter_mean': 0.7657922506332397, 'reward_meter_std': 0.3170587420463562, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9681443572044373, 'reward_repeat_penalty_std': 0.040724486112594604, 'reward_near_duplicate_penalty_mean': 0.9778196811676025, 'reward_near_duplicate_penalty_std': 0.020623713731765747, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9898785352706909, 'reward_distinct_2_std': 0.028627805411815643, 'reward_judge_quality_mean': 0.5262500047683716, 'reward_judge_quality_std': 0.3414648175239563, 'reward_total_composite_mean': 0.4392695426940918, 'reward_total_composite_std': 0.35829058289527893, 'epoch': 0.04}

 28%|██▊       | 919/3300 [3:41:54<6:10:36,  9.34s/it]INFO 04-12 14:28:06 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 920/3300 [3:42:08<7:09:33, 10.83s/it]2026-04-12 14:28:20,848 | INFO | train_grpo_train | metrics_logged mode=train step=920

                                                      
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 7.215151515151516e-06, 'num_tokens': 1873699.0, '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': 0.4964003264904022, 'rewards/meter/std': 0.3180677592754364, 'rewards/count_adherence/mean': 0.5961538553237915, 'rewards/count_adherence/std': 0.11446520686149597, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.08750000596046448, 'rewards/judge_quality/std': 0.051754921674728394, 'rewards/repeat_penalty/mean': 0.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 0.38417744636535645, 'rewards/near_duplicate_penalty/std': 0.19425716996192932, 'rewards/opening_diversity/mean': 0.676510214805603, 'rewards/opening_diversity/std': 0.2780837416648865, 'rewards/distinct_2/mean': 0.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.03293844312429428, 'rewards/total_composite/std': 0.03485986962914467, 'reward': 0.03293844312429428, 'reward_std': 0.03485986962914467, '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': 0.03293844312429428, 'reward_meter_mean': 0.4964003264904022, 'reward_meter_std': 0.3180677592754364, 'reward_count_adherence_mean': 0.5961538553237915, 'reward_count_adherence_std': 0.11446520686149597, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.0, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 0.38417744636535645, 'reward_near_duplicate_penalty_std': 0.19425716996192932, 'reward_opening_diversity_mean': 0.676510214805603, 'reward_opening_diversity_std': 0.2780837416648865, 'reward_distinct_2_mean': 0.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.08750000596046448, 'reward_judge_quality_std': 0.051754921674728394, 'reward_total_composite_mean': 0.03293844312429428, 'reward_total_composite_std': 0.03485986962914467, 'epoch': 0.04}

 28%|██▊       | 920/3300 [3:42:08<7:09:33, 10.83s/it]INFO 04-12 14:28:21 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 921/3300 [3:42:21<7:31:48, 11.39s/it]2026-04-12 14:28:33,561 | INFO | train_grpo_train | metrics_logged mode=train step=921

                                                      
{'loss': -0.0282, 'grad_norm': 5.820071697235107, 'learning_rate': 7.212121212121212e-06, 'num_tokens': 1875251.0, 'completions/mean_length': 114.0, 'completions/min_length': 51.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 57.142860412597656, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.39822918176651, 'rewards/meter/std': 0.30069607496261597, '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/lexical_plausibility/mean': 0.9511978030204773, 'rewards/lexical_plausibility/std': 0.1380334198474884, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.76244056224823, 'rewards/repeat_penalty/std': 0.18652881681919098, 'rewards/near_duplicate_penalty/mean': 0.9035453796386719, 'rewards/near_duplicate_penalty/std': 0.06584017723798752, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.8878251314163208, 'rewards/distinct_2/std': 0.13968020677566528, 'rewards/total_composite/mean': 0.09247905761003494, 'rewards/total_composite/std': 0.11585243046283722, 'reward': 0.09247905761003494, 'reward_std': 0.11585242301225662, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12907761335372925, 'sampling/sampling_logp_difference/max': 2.7132649421691895, 'sampling/importance_sampling_ratio/min': 0.06631992757320404, 'sampling/importance_sampling_ratio/mean': 1.0065613985061646, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6461570747196674, 'clip_ratio/low_mean': 0.07122299261391163, 'clip_ratio/low_min': 0.07122299261391163, 'clip_ratio/high_mean': 0.0369627233594656, 'clip_ratio/high_max': 0.0369627233594656, 'clip_ratio/region_mean': 0.10818571597337723, 'reward_total_mean': 0.09247905761003494, 'reward_meter_mean': 0.39822918176651, 'reward_meter_std': 0.30069607496261597, '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_lexical_plausibility_mean': 0.9511978030204773, 'reward_lexical_plausibility_std': 0.1380334198474884, 'reward_repeat_penalty_mean': 0.76244056224823, 'reward_repeat_penalty_std': 0.18652881681919098, 'reward_near_duplicate_penalty_mean': 0.9035453796386719, 'reward_near_duplicate_penalty_std': 0.06584017723798752, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.8878251314163208, 'reward_distinct_2_std': 0.13968020677566528, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.09247905761003494, 'reward_total_composite_std': 0.11585243046283722, 'epoch': 0.04}

 28%|██▊       | 921/3300 [3:42:21<7:31:48, 11.39s/it]INFO 04-12 14:28:33 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 922/3300 [3:42:33<7:33:26, 11.44s/it]2026-04-12 14:28:45,109 | INFO | train_grpo_train | metrics_logged mode=train step=922

                                                      
{'loss': 0.0752, 'grad_norm': 5.726091384887695, 'learning_rate': 7.2090909090909104e-06, 'num_tokens': 1878305.0, 'completions/mean_length': 198.75, 'completions/min_length': 162.0, 'completions/max_length': 260.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 198.75, 'completions/min_terminated_length': 162.0, 'completions/max_terminated_length': 260.0, 'rewards/meter/mean': 0.035874683409929276, 'rewards/meter/std': 0.03196307644248009, '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/lexical_plausibility/mean': 0.9983552694320679, 'rewards/lexical_plausibility/std': 0.004652021918445826, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.49264419078826904, 'rewards/repeat_penalty/std': 0.24473358690738678, 'rewards/near_duplicate_penalty/mean': 0.8307230472564697, 'rewards/near_duplicate_penalty/std': 0.10259245336055756, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.5891425609588623, 'rewards/distinct_2/std': 0.23500944674015045, 'rewards/total_composite/mean': 0.009451591409742832, 'rewards/total_composite/std': 0.011937719769775867, 'reward': 0.009451591409742832, 'reward_std': 0.011937718838453293, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.108192577958107, 'sampling/sampling_logp_difference/max': 3.4367706775665283, 'sampling/importance_sampling_ratio/min': 0.032168399542570114, 'sampling/importance_sampling_ratio/mean': 1.00033700466156, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.41789066791534424, 'clip_ratio/low_mean': 0.05300997104495764, 'clip_ratio/low_min': 0.05300997104495764, 'clip_ratio/high_mean': 0.04982141498476267, 'clip_ratio/high_max': 0.04982141498476267, 'clip_ratio/region_mean': 0.1028313860297203, 'reward_total_mean': 0.009451591409742832, 'reward_meter_mean': 0.035874683409929276, 'reward_meter_std': 0.03196307644248009, '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_lexical_plausibility_mean': 0.9983552694320679, 'reward_lexical_plausibility_std': 0.004652021918445826, 'reward_repeat_penalty_mean': 0.49264419078826904, 'reward_repeat_penalty_std': 0.24473358690738678, 'reward_near_duplicate_penalty_mean': 0.8307230472564697, 'reward_near_duplicate_penalty_std': 0.10259245336055756, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.5891425609588623, 'reward_distinct_2_std': 0.23500944674015045, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.009451591409742832, 'reward_total_composite_std': 0.011937719769775867, 'epoch': 0.04}

 28%|██▊       | 922/3300 [3:42:33<7:33:26, 11.44s/it]INFO 04-12 14:28:45 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 923/3300 [3:42:40<6:43:03, 10.17s/it]2026-04-12 14:28:52,327 | INFO | train_grpo_train | metrics_logged mode=train step=923

                                                      
{'loss': 0.0251, 'grad_norm': 12.965958595275879, 'learning_rate': 7.206060606060606e-06, 'num_tokens': 1879740.0, 'completions/mean_length': 28.375, 'completions/min_length': 25.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.375, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.47742098569869995, 'rewards/meter/std': 0.28046905994415283, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6850000023841858, 'rewards/judge_quality/std': 0.2512255907058716, 'rewards/repeat_penalty/mean': 0.9661953449249268, 'rewards/repeat_penalty/std': 0.02142074704170227, 'rewards/near_duplicate_penalty/mean': 0.9661953449249268, 'rewards/near_duplicate_penalty/std': 0.02142074704170227, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3213234543800354, 'rewards/total_composite/std': 0.25986388325691223, 'reward': 0.3213234543800354, 'reward_std': 0.25986388325691223, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10668561607599258, 'sampling/sampling_logp_difference/max': 1.1158335208892822, 'sampling/importance_sampling_ratio/min': 0.3276420831680298, 'sampling/importance_sampling_ratio/mean': 0.9958668351173401, 'sampling/importance_sampling_ratio/max': 1.781607985496521, 'entropy': 0.5035784058272839, 'clip_ratio/low_mean': 0.05851545138284564, 'clip_ratio/low_min': 0.05851545138284564, 'clip_ratio/high_mean': 0.03166666720062494, 'clip_ratio/high_max': 0.03166666720062494, 'clip_ratio/region_mean': 0.09018211858347058, 'reward_total_mean': 0.3213234543800354, 'reward_meter_mean': 0.47742098569869995, 'reward_meter_std': 0.28046905994415283, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9661953449249268, 'reward_repeat_penalty_std': 0.02142074704170227, 'reward_near_duplicate_penalty_mean': 0.9661953449249268, 'reward_near_duplicate_penalty_std': 0.02142074704170227, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6850000023841858, 'reward_judge_quality_std': 0.2512255907058716, 'reward_total_composite_mean': 0.3213234543800354, 'reward_total_composite_std': 0.25986388325691223, 'epoch': 0.04}

 28%|██▊       | 923/3300 [3:42:40<6:43:03, 10.17s/it]INFO 04-12 14:28:52 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 924/3300 [3:42:48<6:18:24,  9.56s/it]2026-04-12 14:29:00,440 | INFO | train_grpo_train | metrics_logged mode=train step=924

                                                      
{'loss': -0.0666, 'grad_norm': 12.38703727722168, 'learning_rate': 7.203030303030304e-06, 'num_tokens': 1881511.0, 'completions/mean_length': 55.375, 'completions/min_length': 13.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.375, 'completions/min_terminated_length': 13.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.5948243141174316, 'rewards/meter/std': 0.3165298402309418, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.2357022762298584, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3712500035762787, 'rewards/judge_quality/std': 0.25670650601387024, 'rewards/repeat_penalty/mean': 0.9577116966247559, 'rewards/repeat_penalty/std': 0.08491834253072739, 'rewards/near_duplicate_penalty/mean': 0.9889616966247559, 'rewards/near_duplicate_penalty/std': 0.013677472248673439, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.23747003078460693, 'rewards/total_composite/std': 0.28637459874153137, 'reward': 0.23747003078460693, 'reward_std': 0.28637459874153137, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14392393827438354, 'sampling/sampling_logp_difference/max': 2.947516918182373, 'sampling/importance_sampling_ratio/min': 0.052469830960035324, 'sampling/importance_sampling_ratio/mean': 0.9821318984031677, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7739315181970596, 'clip_ratio/low_mean': 0.08103625569492579, 'clip_ratio/low_min': 0.08103625569492579, 'clip_ratio/high_mean': 0.047786458395421505, 'clip_ratio/high_max': 0.047786458395421505, 'clip_ratio/region_mean': 0.1288227140903473, 'reward_total_mean': 0.23747003078460693, 'reward_meter_mean': 0.5948243141174316, 'reward_meter_std': 0.3165298402309418, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.2357022762298584, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9577116966247559, 'reward_repeat_penalty_std': 0.08491834253072739, 'reward_near_duplicate_penalty_mean': 0.9889616966247559, 'reward_near_duplicate_penalty_std': 0.013677472248673439, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3712500035762787, 'reward_judge_quality_std': 0.25670650601387024, 'reward_total_composite_mean': 0.23747003078460693, 'reward_total_composite_std': 0.28637459874153137, 'epoch': 0.04}

 28%|██▊       | 924/3300 [3:42:48<6:18:24,  9.56s/it]INFO 04-12 14:29:00 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 925/3300 [3:42:56<5:55:51,  8.99s/it]2026-04-12 14:29:08,111 | INFO | train_grpo_train | metrics_logged mode=train step=925

                                                      
{'loss': 0.0148, 'grad_norm': 14.689720153808594, 'learning_rate': 7.2000000000000005e-06, 'num_tokens': 1883157.0, 'completions/mean_length': 34.75, 'completions/min_length': 29.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.75, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.6108083724975586, 'rewards/meter/std': 0.3183428943157196, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.706250011920929, 'rewards/judge_quality/std': 0.29980650544166565, 'rewards/repeat_penalty/mean': 0.7649840116500854, 'rewards/repeat_penalty/std': 0.1884387731552124, 'rewards/near_duplicate_penalty/mean': 0.890425980091095, 'rewards/near_duplicate_penalty/std': 0.06859797984361649, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9339250326156616, 'rewards/distinct_2/std': 0.05497466400265694, 'rewards/total_composite/mean': 0.439907044172287, 'rewards/total_composite/std': 0.3373517096042633, 'reward': 0.439907044172287, 'reward_std': 0.3373516798019409, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11248049885034561, 'sampling/sampling_logp_difference/max': 2.1725354194641113, 'sampling/importance_sampling_ratio/min': 0.11388850212097168, 'sampling/importance_sampling_ratio/mean': 0.9933939576148987, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.37347858399152756, 'clip_ratio/low_mean': 0.035378019558265805, 'clip_ratio/low_min': 0.035378019558265805, 'clip_ratio/high_mean': 0.021780303446576, 'clip_ratio/high_max': 0.021780303446576, 'clip_ratio/region_mean': 0.057158323004841805, 'reward_total_mean': 0.439907044172287, 'reward_meter_mean': 0.6108083724975586, 'reward_meter_std': 0.3183428943157196, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7649840116500854, 'reward_repeat_penalty_std': 0.1884387731552124, 'reward_near_duplicate_penalty_mean': 0.890425980091095, 'reward_near_duplicate_penalty_std': 0.06859797984361649, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9339250326156616, 'reward_distinct_2_std': 0.05497466400265694, 'reward_judge_quality_mean': 0.706250011920929, 'reward_judge_quality_std': 0.29980650544166565, 'reward_total_composite_mean': 0.439907044172287, 'reward_total_composite_std': 0.3373517096042633, 'epoch': 0.04}

 28%|██▊       | 925/3300 [3:42:56<5:55:51,  8.99s/it]INFO 04-12 14:29:08 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 926/3300 [3:43:04<5:47:04,  8.77s/it]2026-04-12 14:29:16,386 | INFO | train_grpo_train | metrics_logged mode=train step=926

                                                      
{'loss': 0.0084, 'grad_norm': 10.720853805541992, 'learning_rate': 7.196969696969698e-06, 'num_tokens': 1884831.0, 'completions/mean_length': 51.25, 'completions/min_length': 49.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.25, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.7739719152450562, 'rewards/meter/std': 0.208439439535141, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4462500214576721, 'rewards/judge_quality/std': 0.19668231904506683, 'rewards/repeat_penalty/mean': 0.7172994017601013, 'rewards/repeat_penalty/std': 0.17979948222637177, 'rewards/near_duplicate_penalty/mean': 0.8759824633598328, 'rewards/near_duplicate_penalty/std': 0.0668049231171608, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.8574627041816711, 'rewards/distinct_2/std': 0.07615248113870621, 'rewards/total_composite/mean': 0.35127681493759155, 'rewards/total_composite/std': 0.18588364124298096, 'reward': 0.35127681493759155, 'reward_std': 0.18588364124298096, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0793343037366867, 'sampling/sampling_logp_difference/max': 1.1771677732467651, 'sampling/importance_sampling_ratio/min': 0.3081502616405487, 'sampling/importance_sampling_ratio/mean': 1.0158623456954956, 'sampling/importance_sampling_ratio/max': 1.9797347784042358, 'entropy': 0.4230548366904259, 'clip_ratio/low_mean': 0.044617616571485996, 'clip_ratio/low_min': 0.044617616571485996, 'clip_ratio/high_mean': 0.028976240195333958, 'clip_ratio/high_max': 0.028976240195333958, 'clip_ratio/region_mean': 0.07359385676681995, 'reward_total_mean': 0.35127681493759155, 'reward_meter_mean': 0.7739719152450562, 'reward_meter_std': 0.208439439535141, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7172994017601013, 'reward_repeat_penalty_std': 0.17979948222637177, 'reward_near_duplicate_penalty_mean': 0.8759824633598328, 'reward_near_duplicate_penalty_std': 0.0668049231171608, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.8574627041816711, 'reward_distinct_2_std': 0.07615248113870621, 'reward_judge_quality_mean': 0.4462500214576721, 'reward_judge_quality_std': 0.19668231904506683, 'reward_total_composite_mean': 0.35127681493759155, 'reward_total_composite_std': 0.18588364124298096, 'epoch': 0.04}

 28%|██▊       | 926/3300 [3:43:04<5:47:04,  8.77s/it]INFO 04-12 14:29:16 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 927/3300 [3:43:18<6:52:11, 10.42s/it]2026-04-12 14:29:30,650 | INFO | train_grpo_train | metrics_logged mode=train step=927

                                                      
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 7.193939393939394e-06, 'num_tokens': 1886495.0, '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': 0.4074295163154602, 'rewards/meter/std': 0.39044761657714844, 'rewards/count_adherence/mean': 0.5625, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.125, 'rewards/judge_quality/std': 0.046291008591651917, 'rewards/repeat_penalty/mean': 0.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 0.45676958560943604, 'rewards/near_duplicate_penalty/std': 0.12844498455524445, 'rewards/opening_diversity/mean': 0.5761560201644897, 'rewards/opening_diversity/std': 0.26298144459724426, 'rewards/distinct_2/mean': 0.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.03147437795996666, 'rewards/total_composite/std': 0.03223505616188049, 'reward': 0.03147437795996666, 'reward_std': 0.03223505616188049, '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': 0.03147437795996666, 'reward_meter_mean': 0.4074295163154602, 'reward_meter_std': 0.39044761657714844, 'reward_count_adherence_mean': 0.5625, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.0, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 0.45676958560943604, 'reward_near_duplicate_penalty_std': 0.12844498455524445, 'reward_opening_diversity_mean': 0.5761560201644897, 'reward_opening_diversity_std': 0.26298144459724426, 'reward_distinct_2_mean': 0.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.125, 'reward_judge_quality_std': 0.046291008591651917, 'reward_total_composite_mean': 0.03147437795996666, 'reward_total_composite_std': 0.03223505616188049, 'epoch': 0.04}

 28%|██▊       | 927/3300 [3:43:18<6:52:11, 10.42s/it]INFO 04-12 14:29:30 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 928/3300 [3:43:27<6:38:34, 10.08s/it]2026-04-12 14:29:39,936 | INFO | train_grpo_train | metrics_logged mode=train step=928

                                                      
{'loss': 0.009, 'grad_norm': 7.838740348815918, 'learning_rate': 7.1909090909090914e-06, 'num_tokens': 1888705.0, 'completions/mean_length': 102.25, 'completions/min_length': 92.0, 'completions/max_length': 117.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 102.25, 'completions/min_terminated_length': 92.0, 'completions/max_terminated_length': 117.0, 'rewards/meter/mean': 0.3131006956100464, 'rewards/meter/std': 0.3299437463283539, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.20000000298023224, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.3542599081993103, 'rewards/repeat_penalty/std': 0.2581605017185211, 'rewards/near_duplicate_penalty/mean': 0.7489602565765381, 'rewards/near_duplicate_penalty/std': 0.11391827464103699, 'rewards/opening_diversity/mean': 0.71875, 'rewards/opening_diversity/std': 0.31160587072372437, 'rewards/distinct_2/mean': 0.5697768926620483, 'rewards/distinct_2/std': 0.19903554022312164, 'rewards/total_composite/mean': 0.05202285945415497, 'rewards/total_composite/std': 0.046992819756269455, 'reward': 0.05202285945415497, 'reward_std': 0.046992819756269455, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09348977357149124, 'sampling/sampling_logp_difference/max': 1.4959441423416138, 'sampling/importance_sampling_ratio/min': 0.22403699159622192, 'sampling/importance_sampling_ratio/mean': 1.0101470947265625, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5535705722868443, 'clip_ratio/low_mean': 0.06675353460013866, 'clip_ratio/low_min': 0.06675353460013866, 'clip_ratio/high_mean': 0.030395023990422487, 'clip_ratio/high_max': 0.030395023990422487, 'clip_ratio/region_mean': 0.09714855859056115, 'reward_total_mean': 0.05202285945415497, 'reward_meter_mean': 0.3131006956100464, 'reward_meter_std': 0.3299437463283539, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.3542599081993103, 'reward_repeat_penalty_std': 0.2581605017185211, 'reward_near_duplicate_penalty_mean': 0.7489602565765381, 'reward_near_duplicate_penalty_std': 0.11391827464103699, 'reward_opening_diversity_mean': 0.71875, 'reward_opening_diversity_std': 0.31160587072372437, 'reward_distinct_2_mean': 0.5697768926620483, 'reward_distinct_2_std': 0.19903554022312164, 'reward_judge_quality_mean': 0.20000000298023224, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.05202285945415497, 'reward_total_composite_std': 0.046992819756269455, 'epoch': 0.04}

 28%|██▊       | 928/3300 [3:43:27<6:38:34, 10.08s/it]INFO 04-12 14:29:40 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 929/3300 [3:43:36<6:24:50,  9.74s/it]2026-04-12 14:29:48,873 | INFO | train_grpo_train | metrics_logged mode=train step=929

                                                      
{'loss': -0.008, 'grad_norm': 7.854971408843994, 'learning_rate': 7.187878787878788e-06, 'num_tokens': 1891248.0, 'completions/mean_length': 122.875, 'completions/min_length': 110.0, 'completions/max_length': 130.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 122.875, 'completions/min_terminated_length': 110.0, 'completions/max_terminated_length': 130.0, 'rewards/meter/mean': 0.21983212232589722, 'rewards/meter/std': 0.28047528862953186, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.7960772514343262, 'rewards/repeat_penalty/std': 0.057195764034986496, 'rewards/near_duplicate_penalty/mean': 0.9249674081802368, 'rewards/near_duplicate_penalty/std': 0.01814199425280094, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8601083159446716, 'rewards/distinct_2/std': 0.049564529210329056, 'rewards/total_composite/mean': 0.053110234439373016, 'rewards/total_composite/std': 0.04758154973387718, 'reward': 0.053110234439373016, 'reward_std': 0.04758154973387718, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1227550134062767, 'sampling/sampling_logp_difference/max': 2.042339324951172, 'sampling/importance_sampling_ratio/min': 0.1297248899936676, 'sampling/importance_sampling_ratio/mean': 0.9976509213447571, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5581958740949631, 'clip_ratio/low_mean': 0.05808962322771549, 'clip_ratio/low_min': 0.05808962322771549, 'clip_ratio/high_mean': 0.044660214334726334, 'clip_ratio/high_max': 0.044660214334726334, 'clip_ratio/region_mean': 0.10274983756244183, 'reward_total_mean': 0.053110234439373016, 'reward_meter_mean': 0.21983212232589722, 'reward_meter_std': 0.28047528862953186, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7960772514343262, 'reward_repeat_penalty_std': 0.057195764034986496, 'reward_near_duplicate_penalty_mean': 0.9249674081802368, 'reward_near_duplicate_penalty_std': 0.01814199425280094, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8601083159446716, 'reward_distinct_2_std': 0.049564529210329056, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.053110234439373016, 'reward_total_composite_std': 0.04758154973387718, 'epoch': 0.04}

 28%|██▊       | 929/3300 [3:43:36<6:24:50,  9.74s/it]INFO 04-12 14:29:49 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 930/3300 [3:43:50<7:07:23, 10.82s/it]2026-04-12 14:30:02,220 | INFO | train_grpo_train | metrics_logged mode=train step=930

                                                      
{'loss': -0.0485, 'grad_norm': 2.702893018722534, 'learning_rate': 7.184848484848486e-06, 'num_tokens': 1892603.0, 'completions/mean_length': 154.375, 'completions/min_length': 29.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 35.16666793823242, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.2841529846191406, 'rewards/meter/std': 0.3785208761692047, '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/lexical_plausibility/mean': 0.9299710988998413, 'rewards/lexical_plausibility/std': 0.13694967329502106, 'rewards/judge_quality/mean': 0.5387499928474426, 'rewards/judge_quality/std': 0.3721918761730194, 'rewards/repeat_penalty/mean': 0.9809141159057617, 'rewards/repeat_penalty/std': 0.01658383198082447, 'rewards/near_duplicate_penalty/mean': 0.9809141159057617, 'rewards/near_duplicate_penalty/std': 0.01658383198082447, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.12557420134544373, 'rewards/total_composite/std': 0.14568524062633514, 'reward': 0.12557420134544373, 'reward_std': 0.14568524062633514, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14664356410503387, 'sampling/sampling_logp_difference/max': 1.1234159469604492, 'sampling/importance_sampling_ratio/min': 0.32516714930534363, 'sampling/importance_sampling_ratio/mean': 0.9998031258583069, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6556182652711868, 'clip_ratio/low_mean': 0.06884021684527397, 'clip_ratio/low_min': 0.06884021684527397, 'clip_ratio/high_mean': 0.06287707574665546, 'clip_ratio/high_max': 0.06287707574665546, 'clip_ratio/region_mean': 0.13171729259192944, 'reward_total_mean': 0.12557420134544373, 'reward_meter_mean': 0.2841529846191406, 'reward_meter_std': 0.3785208761692047, '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_lexical_plausibility_mean': 0.9299710988998413, 'reward_lexical_plausibility_std': 0.13694967329502106, 'reward_repeat_penalty_mean': 0.9809141159057617, 'reward_repeat_penalty_std': 0.01658383198082447, 'reward_near_duplicate_penalty_mean': 0.9809141159057617, 'reward_near_duplicate_penalty_std': 0.01658383198082447, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5387499928474426, 'reward_judge_quality_std': 0.3721918761730194, 'reward_total_composite_mean': 0.12557420134544373, 'reward_total_composite_std': 0.14568524062633514, 'epoch': 0.04}

 28%|██▊       | 930/3300 [3:43:50<7:07:23, 10.82s/it]INFO 04-12 14:30:02 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 931/3300 [3:44:03<7:40:44, 11.67s/it]2026-04-12 14:30:15,865 | INFO | train_grpo_train | metrics_logged mode=train step=931

                                                      
{'loss': 0.1002, 'grad_norm': 3.4817070960998535, 'learning_rate': 7.181818181818182e-06, 'num_tokens': 1896890.0, 'completions/mean_length': 379.875, 'completions/min_length': 293.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 361.0000305175781, 'completions/min_terminated_length': 293.0, 'completions/max_terminated_length': 406.0, 'rewards/meter/mean': 0.18033872544765472, 'rewards/meter/std': 0.21552088856697083, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.13197055459022522, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.17500001192092896, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/repeat_penalty/mean': 0.1096273884177208, 'rewards/repeat_penalty/std': 0.1224500760436058, 'rewards/near_duplicate_penalty/mean': 0.6284300088882446, 'rewards/near_duplicate_penalty/std': 0.2059706598520279, 'rewards/opening_diversity/mean': 0.8528807163238525, 'rewards/opening_diversity/std': 0.29160863161087036, 'rewards/distinct_2/mean': 0.1629660725593567, 'rewards/distinct_2/std': 0.17277877032756805, 'rewards/total_composite/mean': 0.014368465170264244, 'rewards/total_composite/std': 0.016890034079551697, 'reward': 0.014368465170264244, 'reward_std': 0.016890034079551697, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06182783842086792, 'sampling/sampling_logp_difference/max': 2.167691230773926, 'sampling/importance_sampling_ratio/min': 0.11444152891635895, 'sampling/importance_sampling_ratio/mean': 0.9986346364021301, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.26006957329809666, 'clip_ratio/low_mean': 0.0391308874823153, 'clip_ratio/low_min': 0.0391308874823153, 'clip_ratio/high_mean': 0.007671734783798456, 'clip_ratio/high_max': 0.007671734783798456, 'clip_ratio/region_mean': 0.04680262226611376, 'reward_total_mean': 0.014368465170264244, 'reward_meter_mean': 0.18033872544765472, 'reward_meter_std': 0.21552088856697083, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.13197055459022522, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.1096273884177208, 'reward_repeat_penalty_std': 0.1224500760436058, 'reward_near_duplicate_penalty_mean': 0.6284300088882446, 'reward_near_duplicate_penalty_std': 0.2059706598520279, 'reward_opening_diversity_mean': 0.8528807163238525, 'reward_opening_diversity_std': 0.29160863161087036, 'reward_distinct_2_mean': 0.1629660725593567, 'reward_distinct_2_std': 0.17277877032756805, 'reward_judge_quality_mean': 0.17500001192092896, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.014368465170264244, 'reward_total_composite_std': 0.016890034079551697, 'epoch': 0.04}

 28%|██▊       | 931/3300 [3:44:03<7:40:44, 11.67s/it]INFO 04-12 14:30:16 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 932/3300 [3:44:12<7:05:47, 10.79s/it]2026-04-12 14:30:24,601 | INFO | train_grpo_train | metrics_logged mode=train step=932

                                                      
{'loss': 0.1226, 'grad_norm': 9.558558464050293, 'learning_rate': 7.17878787878788e-06, 'num_tokens': 1899269.0, 'completions/mean_length': 91.375, 'completions/min_length': 24.0, 'completions/max_length': 120.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 91.375, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 120.0, 'rewards/meter/mean': 0.5181161165237427, 'rewards/meter/std': 0.33923208713531494, 'rewards/count_adherence/mean': 0.84375, 'rewards/count_adherence/std': 0.29693374037742615, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.29625001549720764, 'rewards/judge_quality/std': 0.26773855090141296, 'rewards/repeat_penalty/mean': 0.41542351245880127, 'rewards/repeat_penalty/std': 0.23470863699913025, 'rewards/near_duplicate_penalty/mean': 0.7675788402557373, 'rewards/near_duplicate_penalty/std': 0.12486158311367035, 'rewards/opening_diversity/mean': 0.734375, 'rewards/opening_diversity/std': 0.16952534019947052, 'rewards/distinct_2/mean': 0.6910839080810547, 'rewards/distinct_2/std': 0.23947729170322418, 'rewards/total_composite/mean': 0.09226329624652863, 'rewards/total_composite/std': 0.08412240445613861, 'reward': 0.09226329624652863, 'reward_std': 0.08412240445613861, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11543485522270203, 'sampling/sampling_logp_difference/max': 3.5148420333862305, 'sampling/importance_sampling_ratio/min': 0.029752502217888832, 'sampling/importance_sampling_ratio/mean': 1.0075103044509888, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5495680496096611, 'clip_ratio/low_mean': 0.06268906686455011, 'clip_ratio/low_min': 0.06268906686455011, 'clip_ratio/high_mean': 0.05669376999139786, 'clip_ratio/high_max': 0.05669376999139786, 'clip_ratio/region_mean': 0.11938283685594797, 'reward_total_mean': 0.09226329624652863, 'reward_meter_mean': 0.5181161165237427, 'reward_meter_std': 0.33923208713531494, 'reward_count_adherence_mean': 0.84375, 'reward_count_adherence_std': 0.29693374037742615, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.41542351245880127, 'reward_repeat_penalty_std': 0.23470863699913025, 'reward_near_duplicate_penalty_mean': 0.7675788402557373, 'reward_near_duplicate_penalty_std': 0.12486158311367035, 'reward_opening_diversity_mean': 0.734375, 'reward_opening_diversity_std': 0.16952534019947052, 'reward_distinct_2_mean': 0.6910839080810547, 'reward_distinct_2_std': 0.23947729170322418, 'reward_judge_quality_mean': 0.29625001549720764, 'reward_judge_quality_std': 0.26773855090141296, 'reward_total_composite_mean': 0.09226329624652863, 'reward_total_composite_std': 0.08412240445613861, 'epoch': 0.04}

 28%|██▊       | 932/3300 [3:44:12<7:05:47, 10.79s/it]INFO 04-12 14:30:24 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 933/3300 [3:44:26<7:37:55, 11.61s/it]2026-04-12 14:30:38,119 | INFO | train_grpo_train | metrics_logged mode=train step=933

                                                      
{'loss': -0.0285, 'grad_norm': 1.1851887702941895, 'learning_rate': 7.175757575757576e-06, 'num_tokens': 1900514.0, 'completions/mean_length': 391.625, 'completions/min_length': 29.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.75, 'completions/mean_terminated_length': 30.5, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.12094494700431824, 'rewards/meter/std': 0.14149430394172668, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.25, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.8156816363334656, 'rewards/lexical_plausibility/std': 0.12250631302595139, 'rewards/judge_quality/mean': 0.20875000953674316, 'rewards/judge_quality/std': 0.3196622431278229, 'rewards/repeat_penalty/mean': 0.9620440006256104, 'rewards/repeat_penalty/std': 0.04678796976804733, 'rewards/near_duplicate_penalty/mean': 0.9620440006256104, 'rewards/near_duplicate_penalty/std': 0.04678796976804733, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.014858603477478027, 'rewards/total_composite/std': 0.029011141508817673, 'reward': 0.014858603477478027, 'reward_std': 0.029011141508817673, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1228853389620781, 'sampling/sampling_logp_difference/max': 0.8168175220489502, 'sampling/importance_sampling_ratio/min': 0.44183555245399475, 'sampling/importance_sampling_ratio/mean': 1.0373996496200562, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.21579509973526, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.029364224523305893, 'clip_ratio/high_max': 0.029364224523305893, 'clip_ratio/region_mean': 0.029364224523305893, 'reward_total_mean': 0.014858603477478027, 'reward_meter_mean': 0.12094494700431824, 'reward_meter_std': 0.14149430394172668, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.25, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.8156816363334656, 'reward_lexical_plausibility_std': 0.12250631302595139, 'reward_repeat_penalty_mean': 0.9620440006256104, 'reward_repeat_penalty_std': 0.04678796976804733, 'reward_near_duplicate_penalty_mean': 0.9620440006256104, 'reward_near_duplicate_penalty_std': 0.04678796976804733, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.20875000953674316, 'reward_judge_quality_std': 0.3196622431278229, 'reward_total_composite_mean': 0.014858603477478027, 'reward_total_composite_std': 0.029011141508817673, 'epoch': 0.04}

 28%|██▊       | 933/3300 [3:44:26<7:37:55, 11.61s/it]INFO 04-12 14:30:38 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 934/3300 [3:44:33<6:53:04, 10.48s/it]2026-04-12 14:30:45,952 | INFO | train_grpo_train | metrics_logged mode=train step=934

                                                      
{'loss': 0.0235, 'grad_norm': 14.033300399780273, 'learning_rate': 7.172727272727273e-06, 'num_tokens': 1902114.0, 'completions/mean_length': 38.0, 'completions/min_length': 30.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.0, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.2824785113334656, 'rewards/meter/std': 0.20607887208461761, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6725000143051147, 'rewards/judge_quality/std': 0.2666056752204895, 'rewards/repeat_penalty/mean': 0.9649622440338135, 'rewards/repeat_penalty/std': 0.06368852406740189, 'rewards/near_duplicate_penalty/mean': 0.9727731943130493, 'rewards/near_duplicate_penalty/std': 0.04641183093190193, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9912587404251099, 'rewards/distinct_2/std': 0.024724015966057777, 'rewards/total_composite/mean': 0.1959935575723648, 'rewards/total_composite/std': 0.20038104057312012, 'reward': 0.1959935575723648, 'reward_std': 0.20038104057312012, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10428716242313385, 'sampling/sampling_logp_difference/max': 1.5179781913757324, 'sampling/importance_sampling_ratio/min': 0.21915455162525177, 'sampling/importance_sampling_ratio/mean': 1.0003457069396973, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.47260911017656326, 'clip_ratio/low_mean': 0.06179779674857855, 'clip_ratio/low_min': 0.06179779674857855, 'clip_ratio/high_mean': 0.04129189206287265, 'clip_ratio/high_max': 0.04129189206287265, 'clip_ratio/region_mean': 0.1030896888114512, 'reward_total_mean': 0.1959935575723648, 'reward_meter_mean': 0.2824785113334656, 'reward_meter_std': 0.20607887208461761, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9649622440338135, 'reward_repeat_penalty_std': 0.06368852406740189, 'reward_near_duplicate_penalty_mean': 0.9727731943130493, 'reward_near_duplicate_penalty_std': 0.04641183093190193, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9912587404251099, 'reward_distinct_2_std': 0.024724015966057777, 'reward_judge_quality_mean': 0.6725000143051147, 'reward_judge_quality_std': 0.2666056752204895, 'reward_total_composite_mean': 0.1959935575723648, 'reward_total_composite_std': 0.20038104057312012, 'epoch': 0.04}

 28%|██▊       | 934/3300 [3:44:33<6:53:04, 10.48s/it]INFO 04-12 14:30:46 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 935/3300 [3:44:46<7:22:14, 11.22s/it]2026-04-12 14:30:58,910 | INFO | train_grpo_train | metrics_logged mode=train step=935

                                                      
{'loss': -0.0372, 'grad_norm': 5.022314071655273, 'learning_rate': 7.16969696969697e-06, 'num_tokens': 1903510.0, 'completions/mean_length': 86.5, 'completions/min_length': 23.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 25.71428680419922, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.7247063517570496, 'rewards/meter/std': 0.36471936106681824, '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/lexical_plausibility/mean': 0.9480119943618774, 'rewards/lexical_plausibility/std': 0.07770383358001709, 'rewards/judge_quality/mean': 0.42625001072883606, 'rewards/judge_quality/std': 0.41025906801223755, 'rewards/repeat_penalty/mean': 0.7472584247589111, 'rewards/repeat_penalty/std': 0.005680246744304895, 'rewards/near_duplicate_penalty/mean': 0.9963445663452148, 'rewards/near_duplicate_penalty/std': 0.007573677226901054, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.32386279106140137, 'rewards/total_composite/std': 0.34933874011039734, 'reward': 0.32386279106140137, 'reward_std': 0.34933874011039734, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12835799157619476, 'sampling/sampling_logp_difference/max': 1.215287446975708, 'sampling/importance_sampling_ratio/min': 0.3430015742778778, 'sampling/importance_sampling_ratio/mean': 1.0257527828216553, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.639490369707346, 'clip_ratio/low_mean': 0.07412842195481062, 'clip_ratio/low_min': 0.07412842195481062, 'clip_ratio/high_mean': 0.022241379134356976, 'clip_ratio/high_max': 0.022241379134356976, 'clip_ratio/region_mean': 0.0963698010891676, 'reward_total_mean': 0.32386279106140137, 'reward_meter_mean': 0.7247063517570496, 'reward_meter_std': 0.36471936106681824, '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_lexical_plausibility_mean': 0.9480119943618774, 'reward_lexical_plausibility_std': 0.07770383358001709, 'reward_repeat_penalty_mean': 0.7472584247589111, 'reward_repeat_penalty_std': 0.005680246744304895, 'reward_near_duplicate_penalty_mean': 0.9963445663452148, 'reward_near_duplicate_penalty_std': 0.007573677226901054, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.42625001072883606, 'reward_judge_quality_std': 0.41025906801223755, 'reward_total_composite_mean': 0.32386279106140137, 'reward_total_composite_std': 0.34933874011039734, 'epoch': 0.04}

 28%|██▊       | 935/3300 [3:44:46<7:22:14, 11.22s/it]INFO 04-12 14:30:59 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 936/3300 [3:44:59<7:39:47, 11.67s/it]2026-04-12 14:31:11,632 | INFO | train_grpo_train | metrics_logged mode=train step=936

                                                      
{'loss': -0.043, 'grad_norm': 2.588585138320923, 'learning_rate': 7.166666666666667e-06, 'num_tokens': 1905273.0, 'completions/mean_length': 168.375, 'completions/min_length': 52.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 53.833335876464844, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.17356033623218536, 'rewards/meter/std': 0.26151424646377563, '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/lexical_plausibility/mean': 0.9430640935897827, 'rewards/lexical_plausibility/std': 0.10776101052761078, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.1414213627576828, 'rewards/repeat_penalty/mean': 0.9473590850830078, 'rewards/repeat_penalty/std': 0.09731647372245789, 'rewards/near_duplicate_penalty/mean': 0.9724011421203613, 'rewards/near_duplicate_penalty/std': 0.03113653138279915, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9725275039672852, 'rewards/distinct_2/std': 0.0777040421962738, 'rewards/total_composite/mean': 0.05963197350502014, 'rewards/total_composite/std': 0.09231703728437424, 'reward': 0.05963197350502014, 'reward_std': 0.09231703728437424, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11515209823846817, 'sampling/sampling_logp_difference/max': 2.6831960678100586, 'sampling/importance_sampling_ratio/min': 0.0683443695306778, 'sampling/importance_sampling_ratio/mean': 1.0075757503509521, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.514962762594223, 'clip_ratio/low_mean': 0.05297540035098791, 'clip_ratio/low_min': 0.05297540035098791, 'clip_ratio/high_mean': 0.02350427396595478, 'clip_ratio/high_max': 0.02350427396595478, 'clip_ratio/region_mean': 0.07647967431694269, 'reward_total_mean': 0.05963197350502014, 'reward_meter_mean': 0.17356033623218536, 'reward_meter_std': 0.26151424646377563, '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_lexical_plausibility_mean': 0.9430640935897827, 'reward_lexical_plausibility_std': 0.10776101052761078, 'reward_repeat_penalty_mean': 0.9473590850830078, 'reward_repeat_penalty_std': 0.09731647372245789, 'reward_near_duplicate_penalty_mean': 0.9724011421203613, 'reward_near_duplicate_penalty_std': 0.03113653138279915, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9725275039672852, 'reward_distinct_2_std': 0.0777040421962738, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.1414213627576828, 'reward_total_composite_mean': 0.05963197350502014, 'reward_total_composite_std': 0.09231703728437424, 'epoch': 0.04}

 28%|██▊       | 936/3300 [3:44:59<7:39:47, 11.67s/it]INFO 04-12 14:31:11 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 937/3300 [3:45:12<7:54:30, 12.05s/it]2026-04-12 14:31:24,561 | INFO | train_grpo_train | metrics_logged mode=train step=937

                                                      
{'loss': -0.0473, 'grad_norm': 4.610383033752441, 'learning_rate': 7.163636363636363e-06, 'num_tokens': 1906784.0, 'completions/mean_length': 103.875, 'completions/min_length': 38.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 45.57143020629883, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.7783571481704712, 'rewards/meter/std': 0.2534632682800293, '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/lexical_plausibility/mean': 0.9749683141708374, 'rewards/lexical_plausibility/std': 0.0708003044128418, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.1414213478565216, 'rewards/repeat_penalty/mean': 0.9211511015892029, 'rewards/repeat_penalty/std': 0.07987295091152191, 'rewards/near_duplicate_penalty/mean': 0.9524011015892029, 'rewards/near_duplicate_penalty/std': 0.044352125376462936, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.19262120127677917, 'rewards/total_composite/std': 0.15365070104599, 'reward': 0.19262120127677917, 'reward_std': 0.1536506861448288, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12564300000667572, 'sampling/sampling_logp_difference/max': 1.3577938079833984, 'sampling/importance_sampling_ratio/min': 0.25722765922546387, 'sampling/importance_sampling_ratio/mean': 1.003007173538208, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7043140381574631, 'clip_ratio/low_mean': 0.04213165584951639, 'clip_ratio/low_min': 0.04213165584951639, 'clip_ratio/high_mean': 0.06418352574110031, 'clip_ratio/high_max': 0.06418352574110031, 'clip_ratio/region_mean': 0.1063151815906167, 'reward_total_mean': 0.19262120127677917, 'reward_meter_mean': 0.7783571481704712, 'reward_meter_std': 0.2534632682800293, '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_lexical_plausibility_mean': 0.9749683141708374, 'reward_lexical_plausibility_std': 0.0708003044128418, 'reward_repeat_penalty_mean': 0.9211511015892029, 'reward_repeat_penalty_std': 0.07987295091152191, 'reward_near_duplicate_penalty_mean': 0.9524011015892029, 'reward_near_duplicate_penalty_std': 0.044352125376462936, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.1414213478565216, 'reward_total_composite_mean': 0.19262120127677917, 'reward_total_composite_std': 0.15365070104599, 'epoch': 0.04}

 28%|██▊       | 937/3300 [3:45:12<7:54:30, 12.05s/it]INFO 04-12 14:31:24 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 938/3300 [3:45:27<8:24:43, 12.82s/it]2026-04-12 14:31:39,197 | INFO | train_grpo_train | metrics_logged mode=train step=938

                                                      
{'loss': -0.1297, 'grad_norm': 5.331589698791504, 'learning_rate': 7.1606060606060615e-06, 'num_tokens': 1908851.0, 'completions/mean_length': 154.375, 'completions/min_length': 92.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 103.28572082519531, 'completions/min_terminated_length': 92.0, 'completions/max_terminated_length': 114.0, 'rewards/meter/mean': 0.4976578652858734, 'rewards/meter/std': 0.23789002001285553, '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/lexical_plausibility/mean': 0.9672771692276001, 'rewards/lexical_plausibility/std': 0.09255408495664597, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.12464234977960587, 'rewards/repeat_penalty/mean': 0.4873230457305908, 'rewards/repeat_penalty/std': 0.28735053539276123, 'rewards/near_duplicate_penalty/mean': 0.8268230557441711, 'rewards/near_duplicate_penalty/std': 0.10564229637384415, 'rewards/opening_diversity/mean': 0.7421875, 'rewards/opening_diversity/std': 0.27024605870246887, 'rewards/distinct_2/mean': 0.7422059178352356, 'rewards/distinct_2/std': 0.1854390949010849, 'rewards/total_composite/mean': 0.13141831755638123, 'rewards/total_composite/std': 0.09342443943023682, 'reward': 0.13141831755638123, 'reward_std': 0.09342443943023682, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10175381600856781, 'sampling/sampling_logp_difference/max': 2.6185126304626465, 'sampling/importance_sampling_ratio/min': 0.07291123270988464, 'sampling/importance_sampling_ratio/mean': 1.0056135654449463, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4526893272995949, 'clip_ratio/low_mean': 0.04729529097676277, 'clip_ratio/low_min': 0.04729529097676277, 'clip_ratio/high_mean': 0.04152615135535598, 'clip_ratio/high_max': 0.04152615135535598, 'clip_ratio/region_mean': 0.08882144233211875, 'reward_total_mean': 0.13141831755638123, 'reward_meter_mean': 0.4976578652858734, 'reward_meter_std': 0.23789002001285553, '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_lexical_plausibility_mean': 0.9672771692276001, 'reward_lexical_plausibility_std': 0.09255408495664597, 'reward_repeat_penalty_mean': 0.4873230457305908, 'reward_repeat_penalty_std': 0.28735053539276123, 'reward_near_duplicate_penalty_mean': 0.8268230557441711, 'reward_near_duplicate_penalty_std': 0.10564229637384415, 'reward_opening_diversity_mean': 0.7421875, 'reward_opening_diversity_std': 0.27024605870246887, 'reward_distinct_2_mean': 0.7422059178352356, 'reward_distinct_2_std': 0.1854390949010849, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.12464234977960587, 'reward_total_composite_mean': 0.13141831755638123, 'reward_total_composite_std': 0.09342443943023682, 'epoch': 0.04}

 28%|██▊       | 938/3300 [3:45:27<8:24:43, 12.82s/it]INFO 04-12 14:31:39 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 939/3300 [3:45:34<7:21:50, 11.23s/it]2026-04-12 14:31:46,707 | INFO | train_grpo_train | metrics_logged mode=train step=939

                                                      
{'loss': -0.0013, 'grad_norm': 22.774871826171875, 'learning_rate': 7.157575757575758e-06, 'num_tokens': 1910220.0, 'completions/mean_length': 20.125, 'completions/min_length': 17.0, 'completions/max_length': 23.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.125, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.24795836210250854, 'rewards/meter/std': 0.3311229646205902, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4675000011920929, 'rewards/judge_quality/std': 0.3022179901599884, 'rewards/repeat_penalty/mean': 0.7350332736968994, 'rewards/repeat_penalty/std': 0.028067149221897125, 'rewards/near_duplicate_penalty/mean': 0.9800443649291992, 'rewards/near_duplicate_penalty/std': 0.03742285817861557, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1447468101978302, 'rewards/total_composite/std': 0.24127833545207977, 'reward': 0.1447468101978302, 'reward_std': 0.24127835035324097, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1540123075246811, 'sampling/sampling_logp_difference/max': 1.2895407676696777, 'sampling/importance_sampling_ratio/min': 0.2892768085002899, 'sampling/importance_sampling_ratio/mean': 1.0504430532455444, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8906053602695465, 'clip_ratio/low_mean': 0.06446102540940046, 'clip_ratio/low_min': 0.06446102540940046, 'clip_ratio/high_mean': 0.01682194648310542, 'clip_ratio/high_max': 0.01682194648310542, 'clip_ratio/region_mean': 0.08128297189250588, 'reward_total_mean': 0.1447468101978302, 'reward_meter_mean': 0.24795836210250854, 'reward_meter_std': 0.3311229646205902, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7350332736968994, 'reward_repeat_penalty_std': 0.028067149221897125, 'reward_near_duplicate_penalty_mean': 0.9800443649291992, 'reward_near_duplicate_penalty_std': 0.03742285817861557, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4675000011920929, 'reward_judge_quality_std': 0.3022179901599884, 'reward_total_composite_mean': 0.1447468101978302, 'reward_total_composite_std': 0.24127833545207977, 'epoch': 0.04}

 28%|██▊       | 939/3300 [3:45:34<7:21:50, 11.23s/it]INFO 04-12 14:31:46 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 940/3300 [3:45:48<7:48:02, 11.90s/it]2026-04-12 14:32:00,174 | INFO | train_grpo_train | metrics_logged mode=train step=940

                                                      
{'loss': -0.1101, 'grad_norm': 4.2889556884765625, 'learning_rate': 7.154545454545455e-06, 'num_tokens': 1912190.0, 'completions/mean_length': 130.25, 'completions/min_length': 70.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 75.71428680419922, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.7814749479293823, 'rewards/meter/std': 0.21990275382995605, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9692437052726746, 'rewards/lexical_plausibility/std': 0.08699192106723785, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.3619333505630493, 'rewards/repeat_penalty/std': 0.2543255090713501, 'rewards/near_duplicate_penalty/mean': 0.8193493485450745, 'rewards/near_duplicate_penalty/std': 0.08180122077465057, 'rewards/opening_diversity/mean': 0.550000011920929, 'rewards/opening_diversity/std': 0.23603872954845428, 'rewards/distinct_2/mean': 0.7278532385826111, 'rewards/distinct_2/std': 0.14499174058437347, 'rewards/total_composite/mean': 0.22237220406532288, 'rewards/total_composite/std': 0.13187643885612488, 'reward': 0.22237220406532288, 'reward_std': 0.13187643885612488, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09415134787559509, 'sampling/sampling_logp_difference/max': 1.5611095428466797, 'sampling/importance_sampling_ratio/min': 0.2099030464887619, 'sampling/importance_sampling_ratio/mean': 1.004854679107666, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4102122113108635, 'clip_ratio/low_mean': 0.012989203911274672, 'clip_ratio/low_min': 0.012989203911274672, 'clip_ratio/high_mean': 0.05385058419778943, 'clip_ratio/high_max': 0.05385058419778943, 'clip_ratio/region_mean': 0.0668397881090641, 'reward_total_mean': 0.22237220406532288, 'reward_meter_mean': 0.7814749479293823, 'reward_meter_std': 0.21990275382995605, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9692437052726746, 'reward_lexical_plausibility_std': 0.08699192106723785, 'reward_repeat_penalty_mean': 0.3619333505630493, 'reward_repeat_penalty_std': 0.2543255090713501, 'reward_near_duplicate_penalty_mean': 0.8193493485450745, 'reward_near_duplicate_penalty_std': 0.08180122077465057, 'reward_opening_diversity_mean': 0.550000011920929, 'reward_opening_diversity_std': 0.23603872954845428, 'reward_distinct_2_mean': 0.7278532385826111, 'reward_distinct_2_std': 0.14499174058437347, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.22237220406532288, 'reward_total_composite_std': 0.13187643885612488, 'epoch': 0.04}

 28%|██▊       | 940/3300 [3:45:48<7:48:02, 11.90s/it]INFO 04-12 14:32:00 [block_pool.py:378] Successfully reset prefix cache

 29%|██▊       | 941/3300 [3:45:57<7:12:57, 11.01s/it]2026-04-12 14:32:09,116 | INFO | train_grpo_train | metrics_logged mode=train step=941

                                                      
{'loss': 0.0581, 'grad_norm': 11.848057746887207, 'learning_rate': 7.151515151515152e-06, 'num_tokens': 1913816.0, 'completions/mean_length': 49.25, 'completions/min_length': 38.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.25, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.13591483235359192, 'rewards/meter/std': 0.20749308168888092, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4712499976158142, 'rewards/judge_quality/std': 0.18795421719551086, 'rewards/repeat_penalty/mean': 0.8903137445449829, 'rewards/repeat_penalty/std': 0.10477173328399658, 'rewards/near_duplicate_penalty/mean': 0.9326732754707336, 'rewards/near_duplicate_penalty/std': 0.04961961507797241, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9526861906051636, 'rewards/distinct_2/std': 0.0784371942281723, 'rewards/total_composite/mean': 0.062338780611753464, 'rewards/total_composite/std': 0.0909540131688118, 'reward': 0.062338780611753464, 'reward_std': 0.0909540131688118, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11286266893148422, 'sampling/sampling_logp_difference/max': 1.3823190927505493, 'sampling/importance_sampling_ratio/min': 0.25099578499794006, 'sampling/importance_sampling_ratio/mean': 1.012656569480896, 'sampling/importance_sampling_ratio/max': 1.821945309638977, 'entropy': 0.7676081731915474, 'clip_ratio/low_mean': 0.04830078175291419, 'clip_ratio/low_min': 0.04830078175291419, 'clip_ratio/high_mean': 0.0328914150595665, 'clip_ratio/high_max': 0.0328914150595665, 'clip_ratio/region_mean': 0.08119219681248069, 'reward_total_mean': 0.062338780611753464, 'reward_meter_mean': 0.13591483235359192, 'reward_meter_std': 0.20749308168888092, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8903137445449829, 'reward_repeat_penalty_std': 0.10477173328399658, 'reward_near_duplicate_penalty_mean': 0.9326732754707336, 'reward_near_duplicate_penalty_std': 0.04961961507797241, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9526861906051636, 'reward_distinct_2_std': 0.0784371942281723, 'reward_judge_quality_mean': 0.4712499976158142, 'reward_judge_quality_std': 0.18795421719551086, 'reward_total_composite_mean': 0.062338780611753464, 'reward_total_composite_std': 0.0909540131688118, 'epoch': 0.04}

 29%|██▊       | 941/3300 [3:45:57<7:12:57, 11.01s/it]INFO 04-12 14:32:09 [block_pool.py:378] Successfully reset prefix cache

 29%|██▊       | 942/3300 [3:46:09<7:33:16, 11.53s/it]2026-04-12 14:32:21,866 | INFO | train_grpo_train | metrics_logged mode=train step=942

                                                      
{'loss': -0.0129, 'grad_norm': 4.764551162719727, 'learning_rate': 7.148484848484849e-06, 'num_tokens': 1915174.0, 'completions/mean_length': 87.75, 'completions/min_length': 23.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 27.142858505249023, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.38765954971313477, 'rewards/meter/std': 0.4857954978942871, '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/lexical_plausibility/mean': 0.9662498235702515, 'rewards/lexical_plausibility/std': 0.07353381812572479, 'rewards/judge_quality/mean': 0.7437500357627869, 'rewards/judge_quality/std': 0.3452095687389374, 'rewards/repeat_penalty/mean': 0.7201552391052246, 'rewards/repeat_penalty/std': 0.04834050312638283, 'rewards/near_duplicate_penalty/mean': 0.9602069854736328, 'rewards/near_duplicate_penalty/std': 0.06445399671792984, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.288032203912735, 'rewards/total_composite/std': 0.39269784092903137, 'reward': 0.288032203912735, 'reward_std': 0.392697811126709, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1669926643371582, 'sampling/sampling_logp_difference/max': 1.0322027206420898, 'sampling/importance_sampling_ratio/min': 0.356221467256546, 'sampling/importance_sampling_ratio/mean': 1.010698676109314, 'sampling/importance_sampling_ratio/max': 1.8058912754058838, 'entropy': 0.9434937760233879, 'clip_ratio/low_mean': 0.08472261484712362, 'clip_ratio/low_min': 0.08472261484712362, 'clip_ratio/high_mean': 0.0729166679084301, 'clip_ratio/high_max': 0.0729166679084301, 'clip_ratio/region_mean': 0.15763928275555372, 'reward_total_mean': 0.288032203912735, 'reward_meter_mean': 0.38765954971313477, 'reward_meter_std': 0.4857954978942871, '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_lexical_plausibility_mean': 0.9662498235702515, 'reward_lexical_plausibility_std': 0.07353381812572479, 'reward_repeat_penalty_mean': 0.7201552391052246, 'reward_repeat_penalty_std': 0.04834050312638283, 'reward_near_duplicate_penalty_mean': 0.9602069854736328, 'reward_near_duplicate_penalty_std': 0.06445399671792984, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7437500357627869, 'reward_judge_quality_std': 0.3452095687389374, 'reward_total_composite_mean': 0.288032203912735, 'reward_total_composite_std': 0.39269784092903137, 'epoch': 0.04}

 29%|██▊       | 942/3300 [3:46:09<7:33:16, 11.53s/it]INFO 04-12 14:32:22 [block_pool.py:378] Successfully reset prefix cache

 29%|██▊       | 943/3300 [3:46:23<7:53:26, 12.05s/it]2026-04-12 14:32:35,130 | INFO | train_grpo_train | metrics_logged mode=train step=943

                                                      
{'loss': -0.0233, 'grad_norm': 1.1867685317993164, 'learning_rate': 7.145454545454547e-06, 'num_tokens': 1916600.0, 'completions/mean_length': 207.25, 'completions/min_length': 23.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 24.399999618530273, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.2950335443019867, 'rewards/meter/std': 0.3719969689846039, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.9414190053939819, 'rewards/lexical_plausibility/std': 0.08183559030294418, 'rewards/judge_quality/mean': 0.16249999403953552, 'rewards/judge_quality/std': 0.12464234977960587, 'rewards/repeat_penalty/mean': 0.6808945536613464, 'rewards/repeat_penalty/std': 0.09268177300691605, 'rewards/near_duplicate_penalty/mean': 0.9078593254089355, 'rewards/near_duplicate_penalty/std': 0.12357569485902786, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.06723420321941376, 'rewards/total_composite/std': 0.10936981439590454, 'reward': 0.06723420321941376, 'reward_std': 0.10936981439590454, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13000689446926117, 'sampling/sampling_logp_difference/max': 1.37107515335083, 'sampling/importance_sampling_ratio/min': 0.2538338899612427, 'sampling/importance_sampling_ratio/mean': 1.0033677816390991, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.44338732957839966, 'clip_ratio/low_mean': 0.016304347664117813, 'clip_ratio/low_min': 0.016304347664117813, 'clip_ratio/high_mean': 0.04329085489735007, 'clip_ratio/high_max': 0.04329085489735007, 'clip_ratio/region_mean': 0.059595202561467886, 'reward_total_mean': 0.06723420321941376, 'reward_meter_mean': 0.2950335443019867, 'reward_meter_std': 0.3719969689846039, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.9414190053939819, 'reward_lexical_plausibility_std': 0.08183559030294418, 'reward_repeat_penalty_mean': 0.6808945536613464, 'reward_repeat_penalty_std': 0.09268177300691605, 'reward_near_duplicate_penalty_mean': 0.9078593254089355, 'reward_near_duplicate_penalty_std': 0.12357569485902786, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.16249999403953552, 'reward_judge_quality_std': 0.12464234977960587, 'reward_total_composite_mean': 0.06723420321941376, 'reward_total_composite_std': 0.10936981439590454, 'epoch': 0.04}

 29%|██▊       | 943/3300 [3:46:23<7:53:26, 12.05s/it]INFO 04-12 14:32:35 [block_pool.py:378] Successfully reset prefix cache

 29%|██▊       | 944/3300 [3:46:32<7:17:38, 11.15s/it]2026-04-12 14:32:44,154 | INFO | train_grpo_train | metrics_logged mode=train step=944

                                                      
{'loss': -0.0174, 'grad_norm': 11.05018138885498, 'learning_rate': 7.142424242424243e-06, 'num_tokens': 1918757.0, 'completions/mean_length': 82.625, 'completions/min_length': 70.0, 'completions/max_length': 92.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 82.625, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 92.0, 'rewards/meter/mean': 0.2062428742647171, 'rewards/meter/std': 0.23885111510753632, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/repeat_penalty/mean': 0.798483669757843, 'rewards/repeat_penalty/std': 0.1474352329969406, 'rewards/near_duplicate_penalty/mean': 0.9184349775314331, 'rewards/near_duplicate_penalty/std': 0.05811554938554764, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8801622986793518, 'rewards/distinct_2/std': 0.10319699347019196, 'rewards/total_composite/mean': 0.05402486026287079, 'rewards/total_composite/std': 0.08079516142606735, 'reward': 0.05402486026287079, 'reward_std': 0.08079515397548676, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11469532549381256, 'sampling/sampling_logp_difference/max': 1.8357130289077759, 'sampling/importance_sampling_ratio/min': 0.15949973464012146, 'sampling/importance_sampling_ratio/mean': 0.9925374388694763, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4836905337870121, 'clip_ratio/low_mean': 0.06299840891733766, 'clip_ratio/low_min': 0.06299840891733766, 'clip_ratio/high_mean': 0.03383559547364712, 'clip_ratio/high_max': 0.03383559547364712, 'clip_ratio/region_mean': 0.09683400439098477, 'reward_total_mean': 0.05402486026287079, 'reward_meter_mean': 0.2062428742647171, 'reward_meter_std': 0.23885111510753632, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.798483669757843, 'reward_repeat_penalty_std': 0.1474352329969406, 'reward_near_duplicate_penalty_mean': 0.9184349775314331, 'reward_near_duplicate_penalty_std': 0.05811554938554764, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8801622986793518, 'reward_distinct_2_std': 0.10319699347019196, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.05402486026287079, 'reward_total_composite_std': 0.08079516142606735, 'epoch': 0.04}

 29%|██▊       | 944/3300 [3:46:32<7:17:38, 11.15s/it]INFO 04-12 14:32:44 [block_pool.py:378] Successfully reset prefix cache

 29%|██▊       | 945/3300 [3:46:42<7:04:40, 10.82s/it]2026-04-12 14:32:54,217 | INFO | train_grpo_train | metrics_logged mode=train step=945

                                                      
{'loss': 0.1285, 'grad_norm': 8.566617012023926, 'learning_rate': 7.1393939393939405e-06, 'num_tokens': 1920796.0, 'completions/mean_length': 89.875, 'completions/min_length': 61.0, 'completions/max_length': 114.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 89.875, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 114.0, 'rewards/meter/mean': 0.6206226944923401, 'rewards/meter/std': 0.3615337014198303, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3375000059604645, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.5913224220275879, 'rewards/repeat_penalty/std': 0.2933858633041382, 'rewards/near_duplicate_penalty/mean': 0.870466947555542, 'rewards/near_duplicate_penalty/std': 0.09446787089109421, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.6936204433441162, 'rewards/distinct_2/std': 0.21726767718791962, 'rewards/total_composite/mean': 0.1903325617313385, 'rewards/total_composite/std': 0.11891020834445953, 'reward': 0.1903325617313385, 'reward_std': 0.11891020834445953, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1071522906422615, 'sampling/sampling_logp_difference/max': 1.8659844398498535, 'sampling/importance_sampling_ratio/min': 0.15474380552768707, 'sampling/importance_sampling_ratio/mean': 1.0033358335494995, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5461409725248814, 'clip_ratio/low_mean': 0.05647054733708501, 'clip_ratio/low_min': 0.05647054733708501, 'clip_ratio/high_mean': 0.048155056312680244, 'clip_ratio/high_max': 0.048155056312680244, 'clip_ratio/region_mean': 0.10462560364976525, 'reward_total_mean': 0.1903325617313385, 'reward_meter_mean': 0.6206226944923401, 'reward_meter_std': 0.3615337014198303, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.5913224220275879, 'reward_repeat_penalty_std': 0.2933858633041382, 'reward_near_duplicate_penalty_mean': 0.870466947555542, 'reward_near_duplicate_penalty_std': 0.09446787089109421, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.6936204433441162, 'reward_distinct_2_std': 0.21726767718791962, 'reward_judge_quality_mean': 0.3375000059604645, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.1903325617313385, 'reward_total_composite_std': 0.11891020834445953, 'epoch': 0.04}

 29%|██▊       | 945/3300 [3:46:42<7:04:40, 10.82s/it]INFO 04-12 14:32:54 [block_pool.py:378] Successfully reset prefix cache

 29%|██▊       | 946/3300 [3:46:57<7:53:26, 12.07s/it]2026-04-12 14:33:09,189 | INFO | train_grpo_train | metrics_logged mode=train step=946

                                                      
{'loss': -0.0815, 'grad_norm': 1.8252280950546265, 'learning_rate': 7.136363636363637e-06, 'num_tokens': 1922538.0, 'completions/mean_length': 303.75, 'completions/min_length': 90.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 95.5, 'completions/min_terminated_length': 90.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.19692228734493256, 'rewards/meter/std': 0.29822638630867004, 'rewards/count_adherence/mean': 0.71875, 'rewards/count_adherence/std': 0.33905068039894104, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.8787853717803955, 'rewards/lexical_plausibility/std': 0.1394978016614914, 'rewards/judge_quality/mean': 0.17499999701976776, 'rewards/judge_quality/std': 0.14880476891994476, 'rewards/repeat_penalty/mean': 0.8171951770782471, 'rewards/repeat_penalty/std': 0.18367989361286163, 'rewards/near_duplicate_penalty/mean': 0.9615234136581421, 'rewards/near_duplicate_penalty/std': 0.057910364121198654, 'rewards/opening_diversity/mean': 0.8984375, 'rewards/opening_diversity/std': 0.12472067028284073, 'rewards/distinct_2/mean': 0.9353432059288025, 'rewards/distinct_2/std': 0.10432393848896027, 'rewards/total_composite/mean': 0.06001988798379898, 'rewards/total_composite/std': 0.10479407012462616, 'reward': 0.06001988798379898, 'reward_std': 0.10479407757520676, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09839633852243423, 'sampling/sampling_logp_difference/max': 1.7060518264770508, 'sampling/importance_sampling_ratio/min': 0.1815812885761261, 'sampling/importance_sampling_ratio/mean': 1.0156311988830566, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.305150855332613, 'clip_ratio/low_mean': 0.026577819138765335, 'clip_ratio/low_min': 0.026577819138765335, 'clip_ratio/high_mean': 0.03377659525722265, 'clip_ratio/high_max': 0.03377659525722265, 'clip_ratio/region_mean': 0.06035441439598799, 'reward_total_mean': 0.06001988798379898, 'reward_meter_mean': 0.19692228734493256, 'reward_meter_std': 0.29822638630867004, 'reward_count_adherence_mean': 0.71875, 'reward_count_adherence_std': 0.33905068039894104, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.8787853717803955, 'reward_lexical_plausibility_std': 0.1394978016614914, 'reward_repeat_penalty_mean': 0.8171951770782471, 'reward_repeat_penalty_std': 0.18367989361286163, 'reward_near_duplicate_penalty_mean': 0.9615234136581421, 'reward_near_duplicate_penalty_std': 0.057910364121198654, 'reward_opening_diversity_mean': 0.8984375, 'reward_opening_diversity_std': 0.12472067028284073, 'reward_distinct_2_mean': 0.9353432059288025, 'reward_distinct_2_std': 0.10432393848896027, 'reward_judge_quality_mean': 0.17499999701976776, 'reward_judge_quality_std': 0.14880476891994476, 'reward_total_composite_mean': 0.06001988798379898, 'reward_total_composite_std': 0.10479407012462616, 'epoch': 0.04}

 29%|██▊       | 946/3300 [3:46:57<7:53:26, 12.07s/it]INFO 04-12 14:33:09 [block_pool.py:378] Successfully reset prefix cache

 29%|██▊       | 947/3300 [3:47:11<8:24:45, 12.87s/it]2026-04-12 14:33:23,938 | INFO | train_grpo_train | metrics_logged mode=train step=947

                                                      
{'loss': -0.0656, 'grad_norm': 1.8696402311325073, 'learning_rate': 7.133333333333334e-06, 'num_tokens': 1926035.0, 'completions/mean_length': 356.125, 'completions/min_length': 261.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 304.16668701171875, 'completions/min_terminated_length': 261.0, 'completions/max_terminated_length': 429.0, 'rewards/meter/mean': 0.13978762924671173, 'rewards/meter/std': 0.2086712121963501, 'rewards/count_adherence/mean': 0.6875, 'rewards/count_adherence/std': 0.13363061845302582, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9387198686599731, 'rewards/lexical_plausibility/std': 0.11450379341840744, 'rewards/judge_quality/mean': 0.125, 'rewards/judge_quality/std': 0.046291008591651917, 'rewards/repeat_penalty/mean': 0.12997031211853027, 'rewards/repeat_penalty/std': 0.22362232208251953, 'rewards/near_duplicate_penalty/mean': 0.5348261594772339, 'rewards/near_duplicate_penalty/std': 0.18436315655708313, 'rewards/opening_diversity/mean': 0.6198718547821045, 'rewards/opening_diversity/std': 0.40011414885520935, 'rewards/distinct_2/mean': 0.24475617706775665, 'rewards/distinct_2/std': 0.4018949568271637, 'rewards/total_composite/mean': 0.008057091385126114, 'rewards/total_composite/std': 0.016946634277701378, 'reward': 0.008057091385126114, 'reward_std': 0.016946634277701378, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.03187386691570282, 'sampling/sampling_logp_difference/max': 1.3523850440979004, 'sampling/importance_sampling_ratio/min': 0.2586227059364319, 'sampling/importance_sampling_ratio/mean': 1.003147006034851, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.14334738068282604, 'clip_ratio/low_mean': 0.015329252695664763, 'clip_ratio/low_min': 0.015329252695664763, 'clip_ratio/high_mean': 0.01003355672582984, 'clip_ratio/high_max': 0.01003355672582984, 'clip_ratio/region_mean': 0.025362809421494603, 'reward_total_mean': 0.008057091385126114, 'reward_meter_mean': 0.13978762924671173, 'reward_meter_std': 0.2086712121963501, 'reward_count_adherence_mean': 0.6875, 'reward_count_adherence_std': 0.13363061845302582, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9387198686599731, 'reward_lexical_plausibility_std': 0.11450379341840744, 'reward_repeat_penalty_mean': 0.12997031211853027, 'reward_repeat_penalty_std': 0.22362232208251953, 'reward_near_duplicate_penalty_mean': 0.5348261594772339, 'reward_near_duplicate_penalty_std': 0.18436315655708313, 'reward_opening_diversity_mean': 0.6198718547821045, 'reward_opening_diversity_std': 0.40011414885520935, 'reward_distinct_2_mean': 0.24475617706775665, 'reward_distinct_2_std': 0.4018949568271637, 'reward_judge_quality_mean': 0.125, 'reward_judge_quality_std': 0.046291008591651917, 'reward_total_composite_mean': 0.008057091385126114, 'reward_total_composite_std': 0.016946634277701378, 'epoch': 0.04}

 29%|██▊       | 947/3300 [3:47:11<8:24:45, 12.87s/it]INFO 04-12 14:33:24 [block_pool.py:378] Successfully reset prefix cache

 29%|██▊       | 948/3300 [3:47:26<8:44:19, 13.38s/it]2026-04-12 14:33:38,495 | INFO | train_grpo_train | metrics_logged mode=train step=948

                                                      
{'loss': -0.0581, 'grad_norm': 2.3616604804992676, 'learning_rate': 7.130303030303031e-06, 'num_tokens': 1927508.0, 'completions/mean_length': 220.125, 'completions/min_length': 41.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 45.0, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.4270865023136139, 'rewards/meter/std': 0.3793179988861084, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.927239179611206, 'rewards/lexical_plausibility/std': 0.10967455804347992, 'rewards/judge_quality/mean': 0.4112499952316284, 'rewards/judge_quality/std': 0.36883747577667236, 'rewards/repeat_penalty/mean': 0.8529618382453918, 'rewards/repeat_penalty/std': 0.17410176992416382, 'rewards/near_duplicate_penalty/mean': 0.9479835629463196, 'rewards/near_duplicate_penalty/std': 0.03856028988957405, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9498425722122192, 'rewards/distinct_2/std': 0.07277485728263855, 'rewards/total_composite/mean': 0.23201444745063782, 'rewards/total_composite/std': 0.3258877992630005, 'reward': 0.23201444745063782, 'reward_std': 0.3258877992630005, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11583569645881653, 'sampling/sampling_logp_difference/max': 1.135561466217041, 'sampling/importance_sampling_ratio/min': 0.3562825918197632, 'sampling/importance_sampling_ratio/mean': 1.0094397068023682, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.45324161648750305, 'clip_ratio/low_mean': 0.026259436737746, 'clip_ratio/low_min': 0.026259436737746, 'clip_ratio/high_mean': 0.02660778909921646, 'clip_ratio/high_max': 0.02660778909921646, 'clip_ratio/region_mean': 0.05286722583696246, 'reward_total_mean': 0.23201444745063782, 'reward_meter_mean': 0.4270865023136139, 'reward_meter_std': 0.3793179988861084, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.927239179611206, 'reward_lexical_plausibility_std': 0.10967455804347992, 'reward_repeat_penalty_mean': 0.8529618382453918, 'reward_repeat_penalty_std': 0.17410176992416382, 'reward_near_duplicate_penalty_mean': 0.9479835629463196, 'reward_near_duplicate_penalty_std': 0.03856028988957405, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9498425722122192, 'reward_distinct_2_std': 0.07277485728263855, 'reward_judge_quality_mean': 0.4112499952316284, 'reward_judge_quality_std': 0.36883747577667236, 'reward_total_composite_mean': 0.23201444745063782, 'reward_total_composite_std': 0.3258877992630005, 'epoch': 0.04}

 29%|██▊       | 948/3300 [3:47:26<8:44:19, 13.38s/it]INFO 04-12 14:33:38 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 949/3300 [3:47:39<8:40:26, 13.28s/it]2026-04-12 14:33:51,559 | INFO | train_grpo_train | metrics_logged mode=train step=949

                                                      
{'loss': -0.0584, 'grad_norm': 4.325341701507568, 'learning_rate': 7.127272727272728e-06, 'num_tokens': 1929231.0, 'completions/mean_length': 108.375, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 50.71428680419922, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.3215366303920746, 'rewards/meter/std': 0.3922673463821411, '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/lexical_plausibility/mean': 0.9545540809631348, 'rewards/lexical_plausibility/std': 0.12854047119617462, 'rewards/judge_quality/mean': 0.44624999165534973, 'rewards/judge_quality/std': 0.23627692461013794, 'rewards/repeat_penalty/mean': 0.9176186323165894, 'rewards/repeat_penalty/std': 0.0858609676361084, 'rewards/near_duplicate_penalty/mean': 0.9587317109107971, 'rewards/near_duplicate_penalty/std': 0.04040151461958885, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9893162250518799, 'rewards/distinct_2/std': 0.03021823801100254, 'rewards/total_composite/mean': 0.14552947878837585, 'rewards/total_composite/std': 0.17575044929981232, 'reward': 0.14552947878837585, 'reward_std': 0.17575044929981232, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10548245906829834, 'sampling/sampling_logp_difference/max': 1.2771339416503906, 'sampling/importance_sampling_ratio/min': 0.2788352966308594, 'sampling/importance_sampling_ratio/mean': 1.0009716749191284, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6781614795327187, 'clip_ratio/low_mean': 0.05460147466510534, 'clip_ratio/low_min': 0.05460147466510534, 'clip_ratio/high_mean': 0.048864406533539295, 'clip_ratio/high_max': 0.048864406533539295, 'clip_ratio/region_mean': 0.10346588119864464, 'reward_total_mean': 0.14552947878837585, 'reward_meter_mean': 0.3215366303920746, 'reward_meter_std': 0.3922673463821411, '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_lexical_plausibility_mean': 0.9545540809631348, 'reward_lexical_plausibility_std': 0.12854047119617462, 'reward_repeat_penalty_mean': 0.9176186323165894, 'reward_repeat_penalty_std': 0.0858609676361084, 'reward_near_duplicate_penalty_mean': 0.9587317109107971, 'reward_near_duplicate_penalty_std': 0.04040151461958885, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9893162250518799, 'reward_distinct_2_std': 0.03021823801100254, 'reward_judge_quality_mean': 0.44624999165534973, 'reward_judge_quality_std': 0.23627692461013794, 'reward_total_composite_mean': 0.14552947878837585, 'reward_total_composite_std': 0.17575044929981232, 'epoch': 0.04}

 29%|██▉       | 949/3300 [3:47:39<8:40:26, 13.28s/it]INFO 04-12 14:33:51 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 950/3300 [3:47:52<8:40:24, 13.29s/it]2026-04-12 14:34:04,852 | INFO | train_grpo_train | metrics_logged mode=train step=950

                                                      
{'loss': -0.0819, 'grad_norm': 2.9789159297943115, 'learning_rate': 7.124242424242424e-06, 'num_tokens': 1930938.0, 'completions/mean_length': 177.375, 'completions/min_length': 41.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 65.83333587646484, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.3404233455657959, 'rewards/meter/std': 0.3930603861808777, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9488165378570557, 'rewards/lexical_plausibility/std': 0.09534463286399841, 'rewards/judge_quality/mean': 0.3062499761581421, 'rewards/judge_quality/std': 0.18015369772911072, 'rewards/repeat_penalty/mean': 0.7187220454216003, 'rewards/repeat_penalty/std': 0.18840491771697998, 'rewards/near_duplicate_penalty/mean': 0.921018123626709, 'rewards/near_duplicate_penalty/std': 0.045870207250118256, 'rewards/opening_diversity/mean': 0.84375, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9106107950210571, 'rewards/distinct_2/std': 0.07955474406480789, 'rewards/total_composite/mean': 0.08306921273469925, 'rewards/total_composite/std': 0.1214752122759819, 'reward': 0.08306921273469925, 'reward_std': 0.1214752048254013, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1294136494398117, 'sampling/sampling_logp_difference/max': 1.809103012084961, 'sampling/importance_sampling_ratio/min': 0.16380098462104797, 'sampling/importance_sampling_ratio/mean': 0.987536609172821, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5565838292241096, 'clip_ratio/low_mean': 0.05406598933041096, 'clip_ratio/low_min': 0.05406598933041096, 'clip_ratio/high_mean': 0.03218851424753666, 'clip_ratio/high_max': 0.03218851424753666, 'clip_ratio/region_mean': 0.08625450357794762, 'reward_total_mean': 0.08306921273469925, 'reward_meter_mean': 0.3404233455657959, 'reward_meter_std': 0.3930603861808777, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9488165378570557, 'reward_lexical_plausibility_std': 0.09534463286399841, 'reward_repeat_penalty_mean': 0.7187220454216003, 'reward_repeat_penalty_std': 0.18840491771697998, 'reward_near_duplicate_penalty_mean': 0.921018123626709, 'reward_near_duplicate_penalty_std': 0.045870207250118256, 'reward_opening_diversity_mean': 0.84375, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9106107950210571, 'reward_distinct_2_std': 0.07955474406480789, 'reward_judge_quality_mean': 0.3062499761581421, 'reward_judge_quality_std': 0.18015369772911072, 'reward_total_composite_mean': 0.08306921273469925, 'reward_total_composite_std': 0.1214752122759819, 'epoch': 0.04}

 29%|██▉       | 950/3300 [3:47:52<8:40:24, 13.29s/it]INFO 04-12 14:34:05 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


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

 15%|█▌        | 2/13 [00:11<01:02,  5.71s/it]

 23%|██▎       | 3/13 [00:22<01:19,  7.99s/it]

 31%|███       | 4/13 [00:33<01:23,  9.22s/it]

 38%|███▊      | 5/13 [00:44<01:18,  9.79s/it]

 46%|████▌     | 6/13 [00:56<01:11, 10.27s/it]

 54%|█████▍    | 7/13 [01:07<01:03, 10.52s/it]

 62%|██████▏   | 8/13 [01:19<00:55, 11.18s/it]

 69%|██████▉   | 9/13 [01:31<00:45, 11.36s/it]

 77%|███████▋  | 10/13 [01:45<00:36, 12.33s/it]

 85%|████████▍ | 11/13 [01:57<00:24, 12.16s/it]

 92%|█████████▏| 12/13 [02:09<00:12, 12.01s/it]

100%|██████████| 13/13 [02:20<00:00, 11.70s/it]2026-04-12 14:36:37,445 | INFO | train_grpo_train | metrics_logged mode=eval step=950

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 152.5201, 'eval_samples_per_second': 0.682, 'eval_steps_per_second': 0.085, 'eval_num_tokens': 1930938.0, 'eval_completions/mean_length': 233.10576923076923, 'eval_completions/min_length': 38.38461538461539, 'eval_completions/max_length': 512.0, 'eval_completions/clipped_ratio': 0.25, 'eval_completions/mean_terminated_length': 143.32930931678186, 'eval_completions/min_terminated_length': 38.38461538461539, 'eval_completions/max_terminated_length': 307.0769230769231, 'eval_rewards/meter/mean': 0.32217265035097414, 'eval_rewards/meter/std': 0.2959560568516071, 'eval_rewards/count_adherence/mean': 0.8043372768622178, 'eval_rewards/count_adherence/std': 0.23716447674311125, 'eval_rewards/arabic_clean/mean': 0.8846153846153846, 'eval_rewards/arabic_clean/std': 0.27000388273826015, 'eval_rewards/lexical_plausibility/mean': 0.9768673640031081, 'eval_rewards/lexical_plausibility/std': 0.05574091896414757, 'eval_rewards/judge_quality/mean': 0.3065384649313413, 'eval_rewards/judge_quality/std': 0.2255362610404308, 'eval_rewards/repeat_penalty/mean': 0.53667335326855, 'eval_rewards/repeat_penalty/std': 0.4180158216219682, 'eval_rewards/near_duplicate_penalty/mean': 0.8241668664492093, 'eval_rewards/near_duplicate_penalty/std': 0.19946125493599817, 'eval_rewards/opening_diversity/mean': 0.8744724026093116, 'eval_rewards/opening_diversity/std': 0.1949955070248017, 'eval_rewards/distinct_2/mean': 0.6022984385490417, 'eval_rewards/distinct_2/std': 0.4252288043498993, 'eval_rewards/total_composite/mean': 0.09447776905905741, 'eval_rewards/total_composite/std': 0.13474725057872441, 'eval_reward': 0.09447776905905741, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.03659316138006174, 'eval_sampling/sampling_logp_difference/max': 0.8933232013995831, 'eval_sampling/importance_sampling_ratio/min': 0.41577648428770214, 'eval_sampling/importance_sampling_ratio/mean': 1.0081675694538996, 'eval_sampling/importance_sampling_ratio/max': 1.4964409516407893, 'eval_entropy': 0.3762010381771968, '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': 0.09447776905905741, 'eval_reward_meter_mean': 0.32217265035097414, 'eval_reward_meter_std': 0.2959560568516071, 'eval_reward_count_adherence_mean': 0.8043372768622178, 'eval_reward_count_adherence_std': 0.23716447674311125, 'eval_reward_arabic_clean_mean': 0.8846153846153846, 'eval_reward_arabic_clean_std': 0.27000388273826015, 'eval_reward_lexical_plausibility_mean': 0.9768673640031081, 'eval_reward_lexical_plausibility_std': 0.05574091896414757, 'eval_reward_repeat_penalty_mean': 0.53667335326855, 'eval_reward_repeat_penalty_std': 0.4180158216219682, 'eval_reward_near_duplicate_penalty_mean': 0.8241668664492093, 'eval_reward_near_duplicate_penalty_std': 0.19946125493599817, 'eval_reward_opening_diversity_mean': 0.8744724026093116, 'eval_reward_opening_diversity_std': 0.1949955070248017, 'eval_reward_distinct_2_mean': 0.6022984385490417, 'eval_reward_distinct_2_std': 0.4252288043498993, 'eval_reward_judge_quality_mean': 0.3065384649313413, 'eval_reward_judge_quality_std': 0.2255362610404308, 'eval_reward_total_composite_mean': 0.09447776905905741, 'eval_reward_total_composite_std': 0.13474725057872441, 'epoch': 0.04}

 29%|██▉       | 950/3300 [3:50:25<8:40:24, 13.29s/it]

100%|██████████| 13/13 [02:20<00:00, 11.70s/it]

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

 29%|██▉       | 951/3300 [3:50:42<39:19:19, 60.26s/it]2026-04-12 14:36:54,737 | INFO | train_grpo_train | metrics_logged mode=train step=951

                                                       
{'loss': -0.0131, 'grad_norm': 3.081400156021118, 'learning_rate': 7.121212121212122e-06, 'num_tokens': 1932556.0, 'completions/mean_length': 159.25, 'completions/min_length': 31.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 41.66666793823242, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.26195645332336426, 'rewards/meter/std': 0.2873122990131378, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9349045157432556, 'rewards/lexical_plausibility/std': 0.1212630644440651, 'rewards/judge_quality/mean': 0.35874998569488525, 'rewards/judge_quality/std': 0.27105283737182617, 'rewards/repeat_penalty/mean': 0.9518330097198486, 'rewards/repeat_penalty/std': 0.05202093720436096, 'rewards/near_duplicate_penalty/mean': 0.9617048501968384, 'rewards/near_duplicate_penalty/std': 0.036329396069049835, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9893953800201416, 'rewards/distinct_2/std': 0.028464725241065025, 'rewards/total_composite/mean': 0.11356673389673233, 'rewards/total_composite/std': 0.24205803871154785, 'reward': 0.11356673389673233, 'reward_std': 0.24205805361270905, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12151914089918137, 'sampling/sampling_logp_difference/max': 1.5041049718856812, 'sampling/importance_sampling_ratio/min': 0.22221609950065613, 'sampling/importance_sampling_ratio/mean': 1.0223890542984009, 'sampling/importance_sampling_ratio/max': 1.8228991031646729, 'entropy': 0.6305864825844765, 'clip_ratio/low_mean': 0.08601269917562604, 'clip_ratio/low_min': 0.08601269917562604, 'clip_ratio/high_mean': 0.0243055559694767, 'clip_ratio/high_max': 0.0243055559694767, 'clip_ratio/region_mean': 0.11031825514510274, 'reward_total_mean': 0.11356673389673233, 'reward_meter_mean': 0.26195645332336426, 'reward_meter_std': 0.2873122990131378, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9349045157432556, 'reward_lexical_plausibility_std': 0.1212630644440651, 'reward_repeat_penalty_mean': 0.9518330097198486, 'reward_repeat_penalty_std': 0.05202093720436096, 'reward_near_duplicate_penalty_mean': 0.9617048501968384, 'reward_near_duplicate_penalty_std': 0.036329396069049835, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9893953800201416, 'reward_distinct_2_std': 0.028464725241065025, 'reward_judge_quality_mean': 0.35874998569488525, 'reward_judge_quality_std': 0.27105283737182617, 'reward_total_composite_mean': 0.11356673389673233, 'reward_total_composite_std': 0.24205803871154785, 'epoch': 0.04}

 29%|██▉       | 951/3300 [3:50:42<39:19:19, 60.26s/it]INFO 04-12 14:36:54 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 952/3300 [3:50:59<30:45:12, 47.15s/it]2026-04-12 14:37:11,417 | INFO | train_grpo_train | metrics_logged mode=train step=952

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 7.118181818181819e-06, 'num_tokens': 1934284.0, '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': 0.033693619072437286, 'rewards/meter/std': 0.041362009942531586, 'rewards/count_adherence/mean': 0.8092105388641357, 'rewards/count_adherence/std': 0.26827728748321533, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9755375385284424, 'rewards/lexical_plausibility/std': 0.060537829995155334, 'rewards/judge_quality/mean': 0.13750000298023224, 'rewards/judge_quality/std': 0.0353553406894207, 'rewards/repeat_penalty/mean': 0.11993913352489471, 'rewards/repeat_penalty/std': 0.33923912048339844, 'rewards/near_duplicate_penalty/mean': 0.6990816593170166, 'rewards/near_duplicate_penalty/std': 0.12835505604743958, 'rewards/opening_diversity/mean': 0.8190094232559204, 'rewards/opening_diversity/std': 0.2678840160369873, 'rewards/distinct_2/mean': 0.12396050244569778, 'rewards/distinct_2/std': 0.3506132662296295, 'rewards/total_composite/mean': 0.0028742384165525436, 'rewards/total_composite/std': 0.004440636839717627, 'reward': 0.0028742384165525436, 'reward_std': 0.004440636839717627, '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': 0.0028742384165525436, 'reward_meter_mean': 0.033693619072437286, 'reward_meter_std': 0.041362009942531586, 'reward_count_adherence_mean': 0.8092105388641357, 'reward_count_adherence_std': 0.26827728748321533, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9755375385284424, 'reward_lexical_plausibility_std': 0.060537829995155334, 'reward_repeat_penalty_mean': 0.11993913352489471, 'reward_repeat_penalty_std': 0.33923912048339844, 'reward_near_duplicate_penalty_mean': 0.6990816593170166, 'reward_near_duplicate_penalty_std': 0.12835505604743958, 'reward_opening_diversity_mean': 0.8190094232559204, 'reward_opening_diversity_std': 0.2678840160369873, 'reward_distinct_2_mean': 0.12396050244569778, 'reward_distinct_2_std': 0.3506132662296295, 'reward_judge_quality_mean': 0.13750000298023224, 'reward_judge_quality_std': 0.0353553406894207, 'reward_total_composite_mean': 0.0028742384165525436, 'reward_total_composite_std': 0.004440636839717627, 'epoch': 0.04}

 29%|██▉       | 952/3300 [3:50:59<30:45:12, 47.15s/it]INFO 04-12 14:37:11 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 953/3300 [3:51:08<23:14:50, 35.66s/it]2026-04-12 14:37:20,135 | INFO | train_grpo_train | metrics_logged mode=train step=953

                                                       
{'loss': 0.0394, 'grad_norm': 10.656508445739746, 'learning_rate': 7.115151515151516e-06, 'num_tokens': 1936177.0, 'completions/mean_length': 57.625, 'completions/min_length': 54.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.625, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.0764944851398468, 'rewards/meter/std': 0.07158287614583969, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5550000071525574, 'rewards/judge_quality/std': 0.22790977358818054, 'rewards/repeat_penalty/mean': 0.9644724130630493, 'rewards/repeat_penalty/std': 0.04163263365626335, 'rewards/near_duplicate_penalty/mean': 0.9796984195709229, 'rewards/near_duplicate_penalty/std': 0.02184489369392395, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.98424232006073, 'rewards/distinct_2/std': 0.029250334948301315, 'rewards/total_composite/mean': 0.0417882576584816, 'rewards/total_composite/std': 0.049210358411073685, 'reward': 0.0417882576584816, 'reward_std': 0.049210354685783386, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14806905388832092, 'sampling/sampling_logp_difference/max': 2.5004420280456543, 'sampling/importance_sampling_ratio/min': 0.08204872161149979, 'sampling/importance_sampling_ratio/mean': 1.0075268745422363, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7753765508532524, 'clip_ratio/low_mean': 0.10691748559474945, 'clip_ratio/low_min': 0.10691748559474945, 'clip_ratio/high_mean': 0.026136363856494427, 'clip_ratio/high_max': 0.026136363856494427, 'clip_ratio/region_mean': 0.13305384945124388, 'reward_total_mean': 0.0417882576584816, 'reward_meter_mean': 0.0764944851398468, 'reward_meter_std': 0.07158287614583969, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9644724130630493, 'reward_repeat_penalty_std': 0.04163263365626335, 'reward_near_duplicate_penalty_mean': 0.9796984195709229, 'reward_near_duplicate_penalty_std': 0.02184489369392395, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.98424232006073, 'reward_distinct_2_std': 0.029250334948301315, 'reward_judge_quality_mean': 0.5550000071525574, 'reward_judge_quality_std': 0.22790977358818054, 'reward_total_composite_mean': 0.0417882576584816, 'reward_total_composite_std': 0.049210358411073685, 'epoch': 0.04}

 29%|██▉       | 953/3300 [3:51:08<23:14:50, 35.66s/it]INFO 04-12 14:37:20 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 954/3300 [3:51:16<17:56:38, 27.54s/it]2026-04-12 14:37:28,705 | INFO | train_grpo_train | metrics_logged mode=train step=954

                                                       
{'loss': 0.0514, 'grad_norm': 8.578009605407715, 'learning_rate': 7.1121212121212125e-06, 'num_tokens': 1938150.0, 'completions/mean_length': 68.625, 'completions/min_length': 62.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.625, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.13190773129463196, 'rewards/meter/std': 0.13010066747665405, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.65625, 'rewards/judge_quality/std': 0.3692052662372589, 'rewards/repeat_penalty/mean': 0.9814157485961914, 'rewards/repeat_penalty/std': 0.028654662892222404, 'rewards/near_duplicate_penalty/mean': 0.9889107346534729, 'rewards/near_duplicate_penalty/std': 0.01105500478297472, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9923076629638672, 'rewards/distinct_2/std': 0.021757129579782486, 'rewards/total_composite/mean': 0.11594647169113159, 'rewards/total_composite/std': 0.1248660460114479, 'reward': 0.11594647169113159, 'reward_std': 0.1248660460114479, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1447618156671524, 'sampling/sampling_logp_difference/max': 2.5152437686920166, 'sampling/importance_sampling_ratio/min': 0.08084320276975632, 'sampling/importance_sampling_ratio/mean': 1.0090447664260864, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.793017815798521, 'clip_ratio/low_mean': 0.10794812813401222, 'clip_ratio/low_min': 0.10794812813401222, 'clip_ratio/high_mean': 0.0504148630425334, 'clip_ratio/high_max': 0.0504148630425334, 'clip_ratio/region_mean': 0.15836299117654562, 'reward_total_mean': 0.11594647169113159, 'reward_meter_mean': 0.13190773129463196, 'reward_meter_std': 0.13010066747665405, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9814157485961914, 'reward_repeat_penalty_std': 0.028654662892222404, 'reward_near_duplicate_penalty_mean': 0.9889107346534729, 'reward_near_duplicate_penalty_std': 0.01105500478297472, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9923076629638672, 'reward_distinct_2_std': 0.021757129579782486, 'reward_judge_quality_mean': 0.65625, 'reward_judge_quality_std': 0.3692052662372589, 'reward_total_composite_mean': 0.11594647169113159, 'reward_total_composite_std': 0.1248660460114479, 'epoch': 0.04}

 29%|██▉       | 954/3300 [3:51:16<17:56:38, 27.54s/it]INFO 04-12 14:37:28 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 955/3300 [3:51:25<14:18:46, 21.97s/it]2026-04-12 14:37:37,734 | INFO | train_grpo_train | metrics_logged mode=train step=955

                                                       
{'loss': -0.0061, 'grad_norm': 9.538155555725098, 'learning_rate': 7.10909090909091e-06, 'num_tokens': 1940008.0, 'completions/mean_length': 72.25, 'completions/min_length': 54.0, 'completions/max_length': 82.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.25, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.017059937119483948, 'rewards/meter/std': 0.014070441015064716, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.05345224589109421, 'rewards/repeat_penalty/mean': 0.8316975235939026, 'rewards/repeat_penalty/std': 0.14587804675102234, 'rewards/near_duplicate_penalty/mean': 0.9541003108024597, 'rewards/near_duplicate_penalty/std': 0.04002492502331734, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8881866931915283, 'rewards/distinct_2/std': 0.09018554538488388, 'rewards/total_composite/mean': 0.006688079796731472, 'rewards/total_composite/std': 0.005313948728144169, 'reward': 0.006688079796731472, 'reward_std': 0.005313948728144169, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1290311962366104, 'sampling/sampling_logp_difference/max': 1.8042993545532227, 'sampling/importance_sampling_ratio/min': 0.16458973288536072, 'sampling/importance_sampling_ratio/mean': 1.0029854774475098, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7239727079868317, 'clip_ratio/low_mean': 0.06259495671838522, 'clip_ratio/low_min': 0.06259495671838522, 'clip_ratio/high_mean': 0.07459933403879404, 'clip_ratio/high_max': 0.07459933403879404, 'clip_ratio/region_mean': 0.13719429075717926, 'reward_total_mean': 0.006688079796731472, 'reward_meter_mean': 0.017059937119483948, 'reward_meter_std': 0.014070441015064716, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8316975235939026, 'reward_repeat_penalty_std': 0.14587804675102234, 'reward_near_duplicate_penalty_mean': 0.9541003108024597, 'reward_near_duplicate_penalty_std': 0.04002492502331734, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8881866931915283, 'reward_distinct_2_std': 0.09018554538488388, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.05345224589109421, 'reward_total_composite_mean': 0.006688079796731472, 'reward_total_composite_std': 0.005313948728144169, 'epoch': 0.04}

 29%|██▉       | 955/3300 [3:51:25<14:18:46, 21.97s/it]INFO 04-12 14:37:37 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 956/3300 [3:51:33<11:37:36, 17.86s/it]2026-04-12 14:37:45,999 | INFO | train_grpo_train | metrics_logged mode=train step=956

                                                       
{'loss': -0.042, 'grad_norm': 11.252642631530762, 'learning_rate': 7.106060606060606e-06, 'num_tokens': 1941691.0, 'completions/mean_length': 56.375, 'completions/min_length': 47.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.375, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.4514808654785156, 'rewards/meter/std': 0.3470607101917267, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.17251639068126678, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5512499809265137, 'rewards/judge_quality/std': 0.3072196841239929, 'rewards/repeat_penalty/mean': 0.9487200975418091, 'rewards/repeat_penalty/std': 0.07420414686203003, 'rewards/near_duplicate_penalty/mean': 0.973534882068634, 'rewards/near_duplicate_penalty/std': 0.039218008518218994, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9732379913330078, 'rewards/distinct_2/std': 0.040286533534526825, 'rewards/total_composite/mean': 0.20551379024982452, 'rewards/total_composite/std': 0.16865909099578857, 'reward': 0.20551379024982452, 'reward_std': 0.16865909099578857, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10767313838005066, 'sampling/sampling_logp_difference/max': 1.2054824829101562, 'sampling/importance_sampling_ratio/min': 0.2995474338531494, 'sampling/importance_sampling_ratio/mean': 1.0086873769760132, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6605033278465271, 'clip_ratio/low_mean': 0.058079251088202, 'clip_ratio/low_min': 0.058079251088202, 'clip_ratio/high_mean': 0.052469510585069656, 'clip_ratio/high_max': 0.052469510585069656, 'clip_ratio/region_mean': 0.11054876167327166, 'reward_total_mean': 0.20551379024982452, 'reward_meter_mean': 0.4514808654785156, 'reward_meter_std': 0.3470607101917267, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.17251639068126678, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9487200975418091, 'reward_repeat_penalty_std': 0.07420414686203003, 'reward_near_duplicate_penalty_mean': 0.973534882068634, 'reward_near_duplicate_penalty_std': 0.039218008518218994, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9732379913330078, 'reward_distinct_2_std': 0.040286533534526825, 'reward_judge_quality_mean': 0.5512499809265137, 'reward_judge_quality_std': 0.3072196841239929, 'reward_total_composite_mean': 0.20551379024982452, 'reward_total_composite_std': 0.16865909099578857, 'epoch': 0.04}

 29%|██▉       | 956/3300 [3:51:33<11:37:36, 17.86s/it]INFO 04-12 14:37:46 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 957/3300 [3:51:49<11:07:56, 17.10s/it]2026-04-12 14:38:01,312 | INFO | train_grpo_train | metrics_logged mode=train step=957

                                                       
{'loss': -0.019, 'grad_norm': 5.3322529792785645, 'learning_rate': 7.103030303030304e-06, 'num_tokens': 1943585.0, 'completions/mean_length': 106.75, 'completions/min_length': 45.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 48.85714340209961, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.12140878289937973, 'rewards/meter/std': 0.22122637927532196, '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/lexical_plausibility/mean': 0.976222813129425, 'rewards/lexical_plausibility/std': 0.047351494431495667, 'rewards/judge_quality/mean': 0.4050000309944153, 'rewards/judge_quality/std': 0.33713075518608093, 'rewards/repeat_penalty/mean': 0.8990166187286377, 'rewards/repeat_penalty/std': 0.1547994464635849, 'rewards/near_duplicate_penalty/mean': 0.9588745832443237, 'rewards/near_duplicate_penalty/std': 0.06519302725791931, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9624639749526978, 'rewards/distinct_2/std': 0.07646677643060684, 'rewards/total_composite/mean': 0.04643910378217697, 'rewards/total_composite/std': 0.07654384523630142, 'reward': 0.04643910378217697, 'reward_std': 0.07654384523630142, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12295597046613693, 'sampling/sampling_logp_difference/max': 1.3022866249084473, 'sampling/importance_sampling_ratio/min': 0.3034486174583435, 'sampling/importance_sampling_ratio/mean': 1.009828805923462, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6286008805036545, 'clip_ratio/low_mean': 0.07907897792756557, 'clip_ratio/low_min': 0.07907897792756557, 'clip_ratio/high_mean': 0.0076530613005161285, 'clip_ratio/high_max': 0.0076530613005161285, 'clip_ratio/region_mean': 0.0867320392280817, 'reward_total_mean': 0.04643910378217697, 'reward_meter_mean': 0.12140878289937973, 'reward_meter_std': 0.22122637927532196, '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_lexical_plausibility_mean': 0.976222813129425, 'reward_lexical_plausibility_std': 0.047351494431495667, 'reward_repeat_penalty_mean': 0.8990166187286377, 'reward_repeat_penalty_std': 0.1547994464635849, 'reward_near_duplicate_penalty_mean': 0.9588745832443237, 'reward_near_duplicate_penalty_std': 0.06519302725791931, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9624639749526978, 'reward_distinct_2_std': 0.07646677643060684, 'reward_judge_quality_mean': 0.4050000309944153, 'reward_judge_quality_std': 0.33713075518608093, 'reward_total_composite_mean': 0.04643910378217697, 'reward_total_composite_std': 0.07654384523630142, 'epoch': 0.04}

 29%|██▉       | 957/3300 [3:51:49<11:07:56, 17.10s/it]INFO 04-12 14:38:01 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 958/3300 [3:52:01<10:14:42, 15.75s/it]2026-04-12 14:38:13,924 | INFO | train_grpo_train | metrics_logged mode=train step=958

                                                       
{'loss': -0.0184, 'grad_norm': 3.158085346221924, 'learning_rate': 7.100000000000001e-06, 'num_tokens': 1944944.0, 'completions/mean_length': 147.875, 'completions/min_length': 22.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 26.5, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.012082438915967941, 'rewards/meter/std': 0.01850433647632599, '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/lexical_plausibility/mean': 0.9354345798492432, 'rewards/lexical_plausibility/std': 0.12001435458660126, 'rewards/judge_quality/mean': 0.4424999952316284, 'rewards/judge_quality/std': 0.41427215933799744, 'rewards/repeat_penalty/mean': 0.8417658805847168, 'rewards/repeat_penalty/std': 0.1267436444759369, 'rewards/near_duplicate_penalty/mean': 0.9980158805847168, 'rewards/near_duplicate_penalty/std': 0.00561195844784379, 'rewards/opening_diversity/mean': 0.84375, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.0044240341521799564, 'rewards/total_composite/std': 0.0059194122441112995, 'reward': 0.0044240341521799564, 'reward_std': 0.005919411778450012, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14891193807125092, 'sampling/sampling_logp_difference/max': 1.3504328727722168, 'sampling/importance_sampling_ratio/min': 0.25912806391716003, 'sampling/importance_sampling_ratio/mean': 1.0550377368927002, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9070471301674843, 'clip_ratio/low_mean': 0.045018117409199476, 'clip_ratio/low_min': 0.045018117409199476, 'clip_ratio/high_mean': 0.036203380674123764, 'clip_ratio/high_max': 0.036203380674123764, 'clip_ratio/region_mean': 0.08122149808332324, 'reward_total_mean': 0.0044240341521799564, 'reward_meter_mean': 0.012082438915967941, 'reward_meter_std': 0.01850433647632599, '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_lexical_plausibility_mean': 0.9354345798492432, 'reward_lexical_plausibility_std': 0.12001435458660126, 'reward_repeat_penalty_mean': 0.8417658805847168, 'reward_repeat_penalty_std': 0.1267436444759369, 'reward_near_duplicate_penalty_mean': 0.9980158805847168, 'reward_near_duplicate_penalty_std': 0.00561195844784379, 'reward_opening_diversity_mean': 0.84375, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4424999952316284, 'reward_judge_quality_std': 0.41427215933799744, 'reward_total_composite_mean': 0.0044240341521799564, 'reward_total_composite_std': 0.0059194122441112995, 'epoch': 0.04}

 29%|██▉       | 958/3300 [3:52:01<10:14:42, 15.75s/it]INFO 04-12 14:38:14 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 959/3300 [3:52:13<9:29:59, 14.61s/it] 2026-04-12 14:38:25,853 | INFO | train_grpo_train | metrics_logged mode=train step=959

                                                      
{'loss': 0.0287, 'grad_norm': 3.0545456409454346, 'learning_rate': 7.096969696969698e-06, 'num_tokens': 1948817.0, 'completions/mean_length': 261.125, 'completions/min_length': 223.0, 'completions/max_length': 331.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 261.125, 'completions/min_terminated_length': 223.0, 'completions/max_terminated_length': 331.0, 'rewards/meter/mean': 0.6300451159477234, 'rewards/meter/std': 0.27191361784935, 'rewards/count_adherence/mean': 0.84375, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.17500001192092896, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/repeat_penalty/mean': 0.07159321010112762, 'rewards/repeat_penalty/std': 0.17123988270759583, 'rewards/near_duplicate_penalty/mean': 0.5685834884643555, 'rewards/near_duplicate_penalty/std': 0.2125176340341568, 'rewards/opening_diversity/mean': 0.6187499761581421, 'rewards/opening_diversity/std': 0.3405396640300751, 'rewards/distinct_2/mean': 0.09342575818300247, 'rewards/distinct_2/std': 0.2187979817390442, 'rewards/total_composite/mean': 0.08568412065505981, 'rewards/total_composite/std': 0.03065934032201767, 'reward': 0.08568412065505981, 'reward_std': 0.03065933845937252, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04350416362285614, 'sampling/sampling_logp_difference/max': 1.3846819400787354, 'sampling/importance_sampling_ratio/min': 0.28611454367637634, 'sampling/importance_sampling_ratio/mean': 1.007127285003662, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.23359408043324947, 'clip_ratio/low_mean': 0.015763788716867566, 'clip_ratio/low_min': 0.015763788716867566, 'clip_ratio/high_mean': 0.02359083783812821, 'clip_ratio/high_max': 0.02359083783812821, 'clip_ratio/region_mean': 0.039354626554995775, 'reward_total_mean': 0.08568412065505981, 'reward_meter_mean': 0.6300451159477234, 'reward_meter_std': 0.27191361784935, 'reward_count_adherence_mean': 0.84375, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.07159321010112762, 'reward_repeat_penalty_std': 0.17123988270759583, 'reward_near_duplicate_penalty_mean': 0.5685834884643555, 'reward_near_duplicate_penalty_std': 0.2125176340341568, 'reward_opening_diversity_mean': 0.6187499761581421, 'reward_opening_diversity_std': 0.3405396640300751, 'reward_distinct_2_mean': 0.09342575818300247, 'reward_distinct_2_std': 0.2187979817390442, 'reward_judge_quality_mean': 0.17500001192092896, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.08568412065505981, 'reward_total_composite_std': 0.03065934032201767, 'epoch': 0.04}

 29%|██▉       | 959/3300 [3:52:13<9:29:59, 14.61s/it]INFO 04-12 14:38:26 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 960/3300 [3:52:23<8:37:20, 13.27s/it]2026-04-12 14:38:35,971 | INFO | train_grpo_train | metrics_logged mode=train step=960

                                                      
{'loss': -0.0094, 'grad_norm': 9.392050743103027, 'learning_rate': 7.093939393939394e-06, 'num_tokens': 1951157.0, 'completions/mean_length': 110.5, 'completions/min_length': 86.0, 'completions/max_length': 124.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 110.5, 'completions/min_terminated_length': 86.0, 'completions/max_terminated_length': 124.0, 'rewards/meter/mean': 0.11665292084217072, 'rewards/meter/std': 0.12713173031806946, 'rewards/count_adherence/mean': 0.9249999523162842, 'rewards/count_adherence/std': 0.1035098284482956, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.8226999640464783, 'rewards/repeat_penalty/std': 0.0968579426407814, 'rewards/near_duplicate_penalty/mean': 0.9547067880630493, 'rewards/near_duplicate_penalty/std': 0.027048135176301003, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8600120544433594, 'rewards/distinct_2/std': 0.0804043784737587, 'rewards/total_composite/mean': 0.03970111533999443, 'rewards/total_composite/std': 0.04452955722808838, 'reward': 0.03970111533999443, 'reward_std': 0.04452955722808838, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13223081827163696, 'sampling/sampling_logp_difference/max': 2.8338468074798584, 'sampling/importance_sampling_ratio/min': 0.05878628045320511, 'sampling/importance_sampling_ratio/mean': 1.0135679244995117, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6751179546117783, 'clip_ratio/low_mean': 0.1079489616677165, 'clip_ratio/low_min': 0.1079489616677165, 'clip_ratio/high_mean': 0.02451846096664667, 'clip_ratio/high_max': 0.02451846096664667, 'clip_ratio/region_mean': 0.13246742263436317, 'reward_total_mean': 0.03970111533999443, 'reward_meter_mean': 0.11665292084217072, 'reward_meter_std': 0.12713173031806946, 'reward_count_adherence_mean': 0.9249999523162842, 'reward_count_adherence_std': 0.1035098284482956, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8226999640464783, 'reward_repeat_penalty_std': 0.0968579426407814, 'reward_near_duplicate_penalty_mean': 0.9547067880630493, 'reward_near_duplicate_penalty_std': 0.027048135176301003, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8600120544433594, 'reward_distinct_2_std': 0.0804043784737587, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.03970111533999443, 'reward_total_composite_std': 0.04452955722808838, 'epoch': 0.04}

 29%|██▉       | 960/3300 [3:52:23<8:37:20, 13.27s/it]INFO 04-12 14:38:36 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 961/3300 [3:52:38<8:53:12, 13.68s/it]2026-04-12 14:38:50,771 | INFO | train_grpo_train | metrics_logged mode=train step=961

                                                      
{'loss': -0.0346, 'grad_norm': 1.7038575410842896, 'learning_rate': 7.0909090909090916e-06, 'num_tokens': 1952589.0, 'completions/mean_length': 277.0, 'completions/min_length': 38.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 42.0, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.19607043266296387, 'rewards/meter/std': 0.27791333198547363, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/lexical_plausibility/mean': 0.8551688194274902, 'rewards/lexical_plausibility/std': 0.16290898621082306, 'rewards/judge_quality/mean': 0.23749999701976776, 'rewards/judge_quality/std': 0.2295181304216385, 'rewards/repeat_penalty/mean': 0.989749550819397, 'rewards/repeat_penalty/std': 0.010598878376185894, 'rewards/near_duplicate_penalty/mean': 0.9902413487434387, 'rewards/near_duplicate_penalty/std': 0.011016878299415112, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9995081424713135, 'rewards/distinct_2/std': 0.0013911199057474732, 'rewards/total_composite/mean': 0.06815202534198761, 'rewards/total_composite/std': 0.1331692785024643, 'reward': 0.06815202534198761, 'reward_std': 0.1331692934036255, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12479446083307266, 'sampling/sampling_logp_difference/max': 1.8015488386154175, 'sampling/importance_sampling_ratio/min': 0.16504307091236115, 'sampling/importance_sampling_ratio/mean': 0.995728075504303, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2866937965154648, 'clip_ratio/low_mean': 0.03645499050617218, 'clip_ratio/low_min': 0.03645499050617218, 'clip_ratio/high_mean': 0.012195121496915817, 'clip_ratio/high_max': 0.012195121496915817, 'clip_ratio/region_mean': 0.048650112003088, 'reward_total_mean': 0.06815202534198761, 'reward_meter_mean': 0.19607043266296387, 'reward_meter_std': 0.27791333198547363, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_lexical_plausibility_mean': 0.8551688194274902, 'reward_lexical_plausibility_std': 0.16290898621082306, 'reward_repeat_penalty_mean': 0.989749550819397, 'reward_repeat_penalty_std': 0.010598878376185894, 'reward_near_duplicate_penalty_mean': 0.9902413487434387, 'reward_near_duplicate_penalty_std': 0.011016878299415112, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9995081424713135, 'reward_distinct_2_std': 0.0013911199057474732, 'reward_judge_quality_mean': 0.23749999701976776, 'reward_judge_quality_std': 0.2295181304216385, 'reward_total_composite_mean': 0.06815202534198761, 'reward_total_composite_std': 0.1331692785024643, 'epoch': 0.04}

 29%|██▉       | 961/3300 [3:52:38<8:53:12, 13.68s/it]INFO 04-12 14:38:51 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 962/3300 [3:52:46<7:51:27, 12.10s/it]2026-04-12 14:38:59,059 | INFO | train_grpo_train | metrics_logged mode=train step=962

                                                      
{'loss': 0.0015, 'grad_norm': 14.421672821044922, 'learning_rate': 7.087878787878788e-06, 'num_tokens': 1954227.0, 'completions/mean_length': 38.75, 'completions/min_length': 37.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.75, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.31433728337287903, 'rewards/meter/std': 0.3194902837276459, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8650000095367432, 'rewards/judge_quality/std': 0.16801361739635468, 'rewards/repeat_penalty/mean': 0.9965499639511108, 'rewards/repeat_penalty/std': 0.009758218191564083, 'rewards/near_duplicate_penalty/mean': 0.9965499639511108, 'rewards/near_duplicate_penalty/std': 0.009758218191564083, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2631286382675171, 'rewards/total_composite/std': 0.2899884879589081, 'reward': 0.2631286382675171, 'reward_std': 0.2899884879589081, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1244831308722496, 'sampling/sampling_logp_difference/max': 2.097195625305176, 'sampling/importance_sampling_ratio/min': 0.1228003203868866, 'sampling/importance_sampling_ratio/mean': 1.0032366514205933, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5374925546348095, 'clip_ratio/low_mean': 0.058975385734811425, 'clip_ratio/low_min': 0.058975385734811425, 'clip_ratio/high_mean': 0.035040403716266155, 'clip_ratio/high_max': 0.035040403716266155, 'clip_ratio/region_mean': 0.09401578945107758, 'reward_total_mean': 0.2631286382675171, 'reward_meter_mean': 0.31433728337287903, 'reward_meter_std': 0.3194902837276459, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9965499639511108, 'reward_repeat_penalty_std': 0.009758218191564083, 'reward_near_duplicate_penalty_mean': 0.9965499639511108, 'reward_near_duplicate_penalty_std': 0.009758218191564083, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8650000095367432, 'reward_judge_quality_std': 0.16801361739635468, 'reward_total_composite_mean': 0.2631286382675171, 'reward_total_composite_std': 0.2899884879589081, 'epoch': 0.04}

 29%|██▉       | 962/3300 [3:52:47<7:51:27, 12.10s/it]INFO 04-12 14:38:59 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 963/3300 [3:53:02<8:27:29, 13.03s/it]2026-04-12 14:39:14,223 | INFO | train_grpo_train | metrics_logged mode=train step=963

                                                      
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 7.084848484848485e-06, 'num_tokens': 1955987.0, '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': 0.5691717863082886, 'rewards/meter/std': 0.3553476929664612, 'rewards/count_adherence/mean': 0.680555522441864, 'rewards/count_adherence/std': 0.07120777666568756, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 0.5089887380599976, 'rewards/near_duplicate_penalty/std': 0.16914601624011993, 'rewards/opening_diversity/mean': 0.48341435194015503, 'rewards/opening_diversity/std': 0.33160334825515747, 'rewards/distinct_2/mean': 0.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.05718400701880455, 'rewards/total_composite/std': 0.03486267477273941, 'reward': 0.05718400701880455, 'reward_std': 0.03486267104744911, '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': 0.05718400701880455, 'reward_meter_mean': 0.5691717863082886, 'reward_meter_std': 0.3553476929664612, 'reward_count_adherence_mean': 0.680555522441864, 'reward_count_adherence_std': 0.07120777666568756, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.0, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 0.5089887380599976, 'reward_near_duplicate_penalty_std': 0.16914601624011993, 'reward_opening_diversity_mean': 0.48341435194015503, 'reward_opening_diversity_std': 0.33160334825515747, 'reward_distinct_2_mean': 0.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.05718400701880455, 'reward_total_composite_std': 0.03486267477273941, 'epoch': 0.04}

 29%|██▉       | 963/3300 [3:53:02<8:27:29, 13.03s/it]INFO 04-12 14:39:14 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 964/3300 [3:53:15<8:35:55, 13.25s/it]2026-04-12 14:39:28,065 | INFO | train_grpo_train | metrics_logged mode=train step=964

                                                      
{'loss': -0.1012, 'grad_norm': 4.135992527008057, 'learning_rate': 7.081818181818182e-06, 'num_tokens': 1958828.0, 'completions/mean_length': 224.125, 'completions/min_length': 156.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 183.00001525878906, 'completions/min_terminated_length': 156.0, 'completions/max_terminated_length': 208.0, 'rewards/meter/mean': 0.17350737750530243, 'rewards/meter/std': 0.2247006595134735, 'rewards/count_adherence/mean': 0.8928571939468384, 'rewards/count_adherence/std': 0.14787118136882782, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9627944827079773, 'rewards/lexical_plausibility/std': 0.1052330806851387, 'rewards/judge_quality/mean': 0.1875, 'rewards/judge_quality/std': 0.1060660183429718, 'rewards/repeat_penalty/mean': 0.29925739765167236, 'rewards/repeat_penalty/std': 0.3122396171092987, 'rewards/near_duplicate_penalty/mean': 0.794903039932251, 'rewards/near_duplicate_penalty/std': 0.11381615698337555, 'rewards/opening_diversity/mean': 0.984375, 'rewards/opening_diversity/std': 0.0289318785071373, 'rewards/distinct_2/mean': 0.3779512643814087, 'rewards/distinct_2/std': 0.3361009657382965, 'rewards/total_composite/mean': 0.023813355714082718, 'rewards/total_composite/std': 0.029519781470298767, 'reward': 0.023813355714082718, 'reward_std': 0.029519781470298767, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07010715454816818, 'sampling/sampling_logp_difference/max': 1.66231107711792, 'sampling/importance_sampling_ratio/min': 0.18970005214214325, 'sampling/importance_sampling_ratio/mean': 1.005490779876709, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.323619969189167, 'clip_ratio/low_mean': 0.0417193821631372, 'clip_ratio/low_min': 0.0417193821631372, 'clip_ratio/high_mean': 0.019327538553625345, 'clip_ratio/high_max': 0.019327538553625345, 'clip_ratio/region_mean': 0.06104692071676254, 'reward_total_mean': 0.023813355714082718, 'reward_meter_mean': 0.17350737750530243, 'reward_meter_std': 0.2247006595134735, 'reward_count_adherence_mean': 0.8928571939468384, 'reward_count_adherence_std': 0.14787118136882782, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9627944827079773, 'reward_lexical_plausibility_std': 0.1052330806851387, 'reward_repeat_penalty_mean': 0.29925739765167236, 'reward_repeat_penalty_std': 0.3122396171092987, 'reward_near_duplicate_penalty_mean': 0.794903039932251, 'reward_near_duplicate_penalty_std': 0.11381615698337555, 'reward_opening_diversity_mean': 0.984375, 'reward_opening_diversity_std': 0.0289318785071373, 'reward_distinct_2_mean': 0.3779512643814087, 'reward_distinct_2_std': 0.3361009657382965, 'reward_judge_quality_mean': 0.1875, 'reward_judge_quality_std': 0.1060660183429718, 'reward_total_composite_mean': 0.023813355714082718, 'reward_total_composite_std': 0.029519781470298767, 'epoch': 0.04}

 29%|██▉       | 964/3300 [3:53:16<8:35:55, 13.25s/it]INFO 04-12 14:39:28 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 965/3300 [3:53:29<8:35:33, 13.25s/it]2026-04-12 14:39:41,233 | INFO | train_grpo_train | metrics_logged mode=train step=965

                                                      
{'loss': -0.0575, 'grad_norm': 5.50330114364624, 'learning_rate': 7.07878787878788e-06, 'num_tokens': 1960719.0, 'completions/mean_length': 109.375, 'completions/min_length': 47.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 51.85714340209961, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.1132308766245842, 'rewards/meter/std': 0.16220051050186157, '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/lexical_plausibility/mean': 0.9751843810081482, 'rewards/lexical_plausibility/std': 0.07018919289112091, 'rewards/judge_quality/mean': 0.4024999737739563, 'rewards/judge_quality/std': 0.26868730783462524, 'rewards/repeat_penalty/mean': 0.9917764067649841, 'rewards/repeat_penalty/std': 0.008495367132127285, 'rewards/near_duplicate_penalty/mean': 0.9917764067649841, 'rewards/near_duplicate_penalty/std': 0.008495367132127285, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.035123150795698166, 'rewards/total_composite/std': 0.058510810136795044, 'reward': 0.035123150795698166, 'reward_std': 0.05851081386208534, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11546876281499863, 'sampling/sampling_logp_difference/max': 1.2601149082183838, 'sampling/importance_sampling_ratio/min': 0.2836214303970337, 'sampling/importance_sampling_ratio/mean': 1.0052849054336548, 'sampling/importance_sampling_ratio/max': 1.872101902961731, 'entropy': 0.5949411317706108, 'clip_ratio/low_mean': 0.03809253824874759, 'clip_ratio/low_min': 0.03809253824874759, 'clip_ratio/high_mean': 0.03663793113082647, 'clip_ratio/high_max': 0.03663793113082647, 'clip_ratio/region_mean': 0.07473046937957406, 'reward_total_mean': 0.035123150795698166, 'reward_meter_mean': 0.1132308766245842, 'reward_meter_std': 0.16220051050186157, '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_lexical_plausibility_mean': 0.9751843810081482, 'reward_lexical_plausibility_std': 0.07018919289112091, 'reward_repeat_penalty_mean': 0.9917764067649841, 'reward_repeat_penalty_std': 0.008495367132127285, 'reward_near_duplicate_penalty_mean': 0.9917764067649841, 'reward_near_duplicate_penalty_std': 0.008495367132127285, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4024999737739563, 'reward_judge_quality_std': 0.26868730783462524, 'reward_total_composite_mean': 0.035123150795698166, 'reward_total_composite_std': 0.058510810136795044, 'epoch': 0.04}

 29%|██▉       | 965/3300 [3:53:29<8:35:33, 13.25s/it]INFO 04-12 14:39:41 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 966/3300 [3:53:42<8:36:21, 13.27s/it]2026-04-12 14:39:54,566 | INFO | train_grpo_train | metrics_logged mode=train step=966

                                                      
{'loss': -0.0374, 'grad_norm': 5.198068141937256, 'learning_rate': 7.075757575757576e-06, 'num_tokens': 1962238.0, 'completions/mean_length': 92.875, 'completions/min_length': 31.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 33.0, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.696131706237793, 'rewards/meter/std': 0.3923903703689575, '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/lexical_plausibility/mean': 0.9717143774032593, 'rewards/lexical_plausibility/std': 0.08000387251377106, 'rewards/judge_quality/mean': 0.35874998569488525, 'rewards/judge_quality/std': 0.27105286717414856, 'rewards/repeat_penalty/mean': 0.7120926380157471, 'rewards/repeat_penalty/std': 0.3322259783744812, 'rewards/near_duplicate_penalty/mean': 0.9344812631607056, 'rewards/near_duplicate_penalty/std': 0.10806850343942642, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.2734534442424774, 'rewards/distinct_2/mean': 0.9171851873397827, 'rewards/distinct_2/std': 0.1039157435297966, 'rewards/total_composite/mean': 0.2924169600009918, 'rewards/total_composite/std': 0.2874559164047241, 'reward': 0.2924169600009918, 'reward_std': 0.2874559164047241, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13471461832523346, 'sampling/sampling_logp_difference/max': 1.5890746116638184, 'sampling/importance_sampling_ratio/min': 0.2041144073009491, 'sampling/importance_sampling_ratio/mean': 1.0088549852371216, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5492743998765945, 'clip_ratio/low_mean': 0.06577036995440722, 'clip_ratio/low_min': 0.06577036995440722, 'clip_ratio/high_mean': 0.0630174521356821, 'clip_ratio/high_max': 0.0630174521356821, 'clip_ratio/region_mean': 0.12878782209008932, 'reward_total_mean': 0.2924169600009918, 'reward_meter_mean': 0.696131706237793, 'reward_meter_std': 0.3923903703689575, '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_lexical_plausibility_mean': 0.9717143774032593, 'reward_lexical_plausibility_std': 0.08000387251377106, 'reward_repeat_penalty_mean': 0.7120926380157471, 'reward_repeat_penalty_std': 0.3322259783744812, 'reward_near_duplicate_penalty_mean': 0.9344812631607056, 'reward_near_duplicate_penalty_std': 0.10806850343942642, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.2734534442424774, 'reward_distinct_2_mean': 0.9171851873397827, 'reward_distinct_2_std': 0.1039157435297966, 'reward_judge_quality_mean': 0.35874998569488525, 'reward_judge_quality_std': 0.27105286717414856, 'reward_total_composite_mean': 0.2924169600009918, 'reward_total_composite_std': 0.2874559164047241, 'epoch': 0.04}

 29%|██▉       | 966/3300 [3:53:42<8:36:21, 13.27s/it]INFO 04-12 14:39:54 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 967/3300 [3:53:52<7:56:18, 12.25s/it]2026-04-12 14:40:04,430 | INFO | train_grpo_train | metrics_logged mode=train step=967

                                                      
{'loss': 0.0089, 'grad_norm': 6.549985885620117, 'learning_rate': 7.072727272727273e-06, 'num_tokens': 1964736.0, 'completions/mean_length': 129.25, 'completions/min_length': 118.0, 'completions/max_length': 139.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 129.25, 'completions/min_terminated_length': 118.0, 'completions/max_terminated_length': 139.0, 'rewards/meter/mean': 0.968479573726654, 'rewards/meter/std': 0.03996254876255989, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.044766947627067566, 'rewards/repeat_penalty/std': 0.02587527595460415, 'rewards/near_duplicate_penalty/mean': 0.5779581665992737, 'rewards/near_duplicate_penalty/std': 0.0901092141866684, 'rewards/opening_diversity/mean': 0.26250001788139343, 'rewards/opening_diversity/std': 0.06943651288747787, 'rewards/distinct_2/mean': 0.2992805242538452, 'rewards/distinct_2/std': 0.11232910305261612, 'rewards/total_composite/mean': 0.33896785974502563, 'rewards/total_composite/std': 0.013986884616315365, 'reward': 0.33896785974502563, 'reward_std': 0.013986876234412193, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0556277297437191, 'sampling/sampling_logp_difference/max': 2.128795623779297, 'sampling/importance_sampling_ratio/min': 0.11898049712181091, 'sampling/importance_sampling_ratio/mean': 1.0075849294662476, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.32943726517260075, 'clip_ratio/low_mean': 0.012544722761958838, 'clip_ratio/low_min': 0.012544722761958838, 'clip_ratio/high_mean': 0.03405781940091401, 'clip_ratio/high_max': 0.03405781940091401, 'clip_ratio/region_mean': 0.04660254216287285, 'reward_total_mean': 0.33896785974502563, 'reward_meter_mean': 0.968479573726654, 'reward_meter_std': 0.03996254876255989, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.044766947627067566, 'reward_repeat_penalty_std': 0.02587527595460415, 'reward_near_duplicate_penalty_mean': 0.5779581665992737, 'reward_near_duplicate_penalty_std': 0.0901092141866684, 'reward_opening_diversity_mean': 0.26250001788139343, 'reward_opening_diversity_std': 0.06943651288747787, 'reward_distinct_2_mean': 0.2992805242538452, 'reward_distinct_2_std': 0.11232910305261612, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.33896785974502563, 'reward_total_composite_std': 0.013986884616315365, 'epoch': 0.04}

 29%|██▉       | 967/3300 [3:53:52<7:56:18, 12.25s/it]INFO 04-12 14:40:04 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 968/3300 [3:54:06<8:17:54, 12.81s/it]2026-04-12 14:40:18,552 | INFO | train_grpo_train | metrics_logged mode=train step=968

                                                      
{'loss': -0.0399, 'grad_norm': 2.4287562370300293, 'learning_rate': 7.06969696969697e-06, 'num_tokens': 1966176.0, 'completions/mean_length': 217.0, 'completions/min_length': 37.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 40.0, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.12095977365970612, 'rewards/meter/std': 0.17009131610393524, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.9163911938667297, 'rewards/lexical_plausibility/std': 0.12047196924686432, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.19955307245254517, 'rewards/repeat_penalty/mean': 0.9134237766265869, 'rewards/repeat_penalty/std': 0.03889590501785278, 'rewards/near_duplicate_penalty/mean': 0.9542394280433655, 'rewards/near_duplicate_penalty/std': 0.04148237407207489, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9579815864562988, 'rewards/distinct_2/std': 0.03885412588715553, 'rewards/total_composite/mean': 0.05122220888733864, 'rewards/total_composite/std': 0.07818522304296494, 'reward': 0.05122220888733864, 'reward_std': 0.07818521559238434, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11332090198993683, 'sampling/sampling_logp_difference/max': 1.26051664352417, 'sampling/importance_sampling_ratio/min': 0.28350749611854553, 'sampling/importance_sampling_ratio/mean': 0.9961549043655396, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3598831035196781, 'clip_ratio/low_mean': 0.03147659823298454, 'clip_ratio/low_min': 0.03147659823298454, 'clip_ratio/high_mean': 0.039835166186094284, 'clip_ratio/high_max': 0.039835166186094284, 'clip_ratio/region_mean': 0.07131176441907883, 'reward_total_mean': 0.05122220888733864, 'reward_meter_mean': 0.12095977365970612, 'reward_meter_std': 0.17009131610393524, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.9163911938667297, 'reward_lexical_plausibility_std': 0.12047196924686432, 'reward_repeat_penalty_mean': 0.9134237766265869, 'reward_repeat_penalty_std': 0.03889590501785278, 'reward_near_duplicate_penalty_mean': 0.9542394280433655, 'reward_near_duplicate_penalty_std': 0.04148237407207489, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9579815864562988, 'reward_distinct_2_std': 0.03885412588715553, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.19955307245254517, 'reward_total_composite_mean': 0.05122220888733864, 'reward_total_composite_std': 0.07818522304296494, 'epoch': 0.04}

 29%|██▉       | 968/3300 [3:54:06<8:17:54, 12.81s/it]INFO 04-12 14:40:18 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 969/3300 [3:54:15<7:30:32, 11.60s/it]2026-04-12 14:40:27,310 | INFO | train_grpo_train | metrics_logged mode=train step=969

                                                      
{'loss': 0.0543, 'grad_norm': 10.847005844116211, 'learning_rate': 7.066666666666667e-06, 'num_tokens': 1968034.0, 'completions/mean_length': 69.25, 'completions/min_length': 60.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.25, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.08024676144123077, 'rewards/meter/std': 0.08288618922233582, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.44999998807907104, 'rewards/judge_quality/std': 0.13093073666095734, 'rewards/repeat_penalty/mean': 0.8615816831588745, 'rewards/repeat_penalty/std': 0.09671135246753693, 'rewards/near_duplicate_penalty/mean': 0.9537315368652344, 'rewards/near_duplicate_penalty/std': 0.022020557895302773, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9020715951919556, 'rewards/distinct_2/std': 0.08497128635644913, 'rewards/total_composite/mean': 0.03415246680378914, 'rewards/total_composite/std': 0.037199120968580246, 'reward': 0.03415246680378914, 'reward_std': 0.037199124693870544, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1389940232038498, 'sampling/sampling_logp_difference/max': 1.6286948919296265, 'sampling/importance_sampling_ratio/min': 0.19618543982505798, 'sampling/importance_sampling_ratio/mean': 1.0234410762786865, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6581175997853279, 'clip_ratio/low_mean': 0.09743837546557188, 'clip_ratio/low_min': 0.09743837546557188, 'clip_ratio/high_mean': 0.03078396897763014, 'clip_ratio/high_max': 0.03078396897763014, 'clip_ratio/region_mean': 0.12822234444320202, 'reward_total_mean': 0.03415246680378914, 'reward_meter_mean': 0.08024676144123077, 'reward_meter_std': 0.08288618922233582, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8615816831588745, 'reward_repeat_penalty_std': 0.09671135246753693, 'reward_near_duplicate_penalty_mean': 0.9537315368652344, 'reward_near_duplicate_penalty_std': 0.022020557895302773, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9020715951919556, 'reward_distinct_2_std': 0.08497128635644913, 'reward_judge_quality_mean': 0.44999998807907104, 'reward_judge_quality_std': 0.13093073666095734, 'reward_total_composite_mean': 0.03415246680378914, 'reward_total_composite_std': 0.037199120968580246, 'epoch': 0.04}

 29%|██▉       | 969/3300 [3:54:15<7:30:32, 11.60s/it]INFO 04-12 14:40:27 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 970/3300 [3:54:23<6:53:44, 10.65s/it]2026-04-12 14:40:35,770 | INFO | train_grpo_train | metrics_logged mode=train step=970

                                                      
{'loss': 0.0095, 'grad_norm': 11.570093154907227, 'learning_rate': 7.063636363636365e-06, 'num_tokens': 1969902.0, 'completions/mean_length': 57.5, 'completions/min_length': 55.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.5, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.690869927406311, 'rewards/meter/std': 0.4130357801914215, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9886363744735718, 'rewards/lexical_plausibility/std': 0.03214120864868164, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.1060660183429718, 'rewards/repeat_penalty/mean': 0.9281680583953857, 'rewards/repeat_penalty/std': 0.08725522458553314, 'rewards/near_duplicate_penalty/mean': 0.9678975939750671, 'rewards/near_duplicate_penalty/std': 0.033893290907144547, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.957433819770813, 'rewards/distinct_2/std': 0.06260645389556885, 'rewards/total_composite/mean': 0.30004560947418213, 'rewards/total_composite/std': 0.18818852305412292, 'reward': 0.30004560947418213, 'reward_std': 0.18818852305412292, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10780466347932816, 'sampling/sampling_logp_difference/max': 1.8387866020202637, 'sampling/importance_sampling_ratio/min': 0.15901026129722595, 'sampling/importance_sampling_ratio/mean': 0.9995007514953613, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5267671011388302, 'clip_ratio/low_mean': 0.029062713496387005, 'clip_ratio/low_min': 0.029062713496387005, 'clip_ratio/high_mean': 0.05143678281456232, 'clip_ratio/high_max': 0.05143678281456232, 'clip_ratio/region_mean': 0.08049949631094933, 'reward_total_mean': 0.30004560947418213, 'reward_meter_mean': 0.690869927406311, 'reward_meter_std': 0.4130357801914215, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9886363744735718, 'reward_lexical_plausibility_std': 0.03214120864868164, 'reward_repeat_penalty_mean': 0.9281680583953857, 'reward_repeat_penalty_std': 0.08725522458553314, 'reward_near_duplicate_penalty_mean': 0.9678975939750671, 'reward_near_duplicate_penalty_std': 0.033893290907144547, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.957433819770813, 'reward_distinct_2_std': 0.06260645389556885, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.1060660183429718, 'reward_total_composite_mean': 0.30004560947418213, 'reward_total_composite_std': 0.18818852305412292, 'epoch': 0.04}

 29%|██▉       | 970/3300 [3:54:23<6:53:44, 10.65s/it]INFO 04-12 14:40:36 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 971/3300 [3:54:33<6:43:01, 10.38s/it]2026-04-12 14:40:45,519 | INFO | train_grpo_train | metrics_logged mode=train step=971

                                                      
{'loss': 0.0161, 'grad_norm': 5.489332675933838, 'learning_rate': 7.060606060606061e-06, 'num_tokens': 1972720.0, 'completions/mean_length': 157.25, 'completions/min_length': 114.0, 'completions/max_length': 172.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 157.25, 'completions/min_terminated_length': 114.0, 'completions/max_terminated_length': 172.0, 'rewards/meter/mean': 0.35954898595809937, 'rewards/meter/std': 0.26351335644721985, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9975490570068359, 'rewards/lexical_plausibility/std': 0.006932419259101152, 'rewards/judge_quality/mean': 0.23750001192092896, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.43385231494903564, 'rewards/repeat_penalty/std': 0.20487482845783234, 'rewards/near_duplicate_penalty/mean': 0.8501541018486023, 'rewards/near_duplicate_penalty/std': 0.06414903700351715, 'rewards/opening_diversity/mean': 0.9140625, 'rewards/opening_diversity/std': 0.16682934761047363, 'rewards/distinct_2/mean': 0.6013368964195251, 'rewards/distinct_2/std': 0.1815468966960907, 'rewards/total_composite/mean': 0.07021139562129974, 'rewards/total_composite/std': 0.05332328751683235, 'reward': 0.07021139562129974, 'reward_std': 0.05332328379154205, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09675942361354828, 'sampling/sampling_logp_difference/max': 2.3524839878082275, 'sampling/importance_sampling_ratio/min': 0.09513255953788757, 'sampling/importance_sampling_ratio/mean': 0.9920095801353455, 'sampling/importance_sampling_ratio/max': 1.8275434970855713, 'entropy': 0.5025036595761776, 'clip_ratio/low_mean': 0.04403558745980263, 'clip_ratio/low_min': 0.04403558745980263, 'clip_ratio/high_mean': 0.03316158428788185, 'clip_ratio/high_max': 0.03316158428788185, 'clip_ratio/region_mean': 0.07719717174768448, 'reward_total_mean': 0.07021139562129974, 'reward_meter_mean': 0.35954898595809937, 'reward_meter_std': 0.26351335644721985, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.09258200973272324, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9975490570068359, 'reward_lexical_plausibility_std': 0.006932419259101152, 'reward_repeat_penalty_mean': 0.43385231494903564, 'reward_repeat_penalty_std': 0.20487482845783234, 'reward_near_duplicate_penalty_mean': 0.8501541018486023, 'reward_near_duplicate_penalty_std': 0.06414903700351715, 'reward_opening_diversity_mean': 0.9140625, 'reward_opening_diversity_std': 0.16682934761047363, 'reward_distinct_2_mean': 0.6013368964195251, 'reward_distinct_2_std': 0.1815468966960907, 'reward_judge_quality_mean': 0.23750001192092896, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.07021139562129974, 'reward_total_composite_std': 0.05332328751683235, 'epoch': 0.04}

 29%|██▉       | 971/3300 [3:54:33<6:43:01, 10.38s/it]INFO 04-12 14:40:45 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 972/3300 [3:54:43<6:38:23, 10.27s/it]2026-04-12 14:40:55,511 | INFO | train_grpo_train | metrics_logged mode=train step=972

                                                      
{'loss': -0.0668, 'grad_norm': 4.7223100662231445, 'learning_rate': 7.057575757575759e-06, 'num_tokens': 1975886.0, 'completions/mean_length': 211.75, 'completions/min_length': 158.0, 'completions/max_length': 268.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 211.75, 'completions/min_terminated_length': 158.0, 'completions/max_terminated_length': 268.0, 'rewards/meter/mean': 0.23917627334594727, 'rewards/meter/std': 0.2844199240207672, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.0578637570142746, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.10000000149011612, 'rewards/judge_quality/std': 0.053452253341674805, 'rewards/repeat_penalty/mean': 8.849838195601478e-05, 'rewards/repeat_penalty/std': 0.0002503112191334367, 'rewards/near_duplicate_penalty/mean': 0.3626028895378113, 'rewards/near_duplicate_penalty/std': 0.24472764134407043, 'rewards/opening_diversity/mean': 0.4099264442920685, 'rewards/opening_diversity/std': 0.3534538447856903, 'rewards/distinct_2/mean': 0.0028586278203874826, 'rewards/distinct_2/std': 0.008085420355200768, 'rewards/total_composite/mean': 0.016795290634036064, 'rewards/total_composite/std': 0.02730741538107395, 'reward': 0.016795290634036064, 'reward_std': 0.027307413518428802, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.036760326474905014, 'sampling/sampling_logp_difference/max': 2.17226243019104, 'sampling/importance_sampling_ratio/min': 0.11391958594322205, 'sampling/importance_sampling_ratio/mean': 1.0065410137176514, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.25723590049892664, 'clip_ratio/low_mean': 0.01981035154312849, 'clip_ratio/low_min': 0.01981035154312849, 'clip_ratio/high_mean': 0.010458126198500395, 'clip_ratio/high_max': 0.010458126198500395, 'clip_ratio/region_mean': 0.030268477741628885, 'reward_total_mean': 0.016795290634036064, 'reward_meter_mean': 0.23917627334594727, 'reward_meter_std': 0.2844199240207672, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.0578637570142746, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 8.849838195601478e-05, 'reward_repeat_penalty_std': 0.0002503112191334367, 'reward_near_duplicate_penalty_mean': 0.3626028895378113, 'reward_near_duplicate_penalty_std': 0.24472764134407043, 'reward_opening_diversity_mean': 0.4099264442920685, 'reward_opening_diversity_std': 0.3534538447856903, 'reward_distinct_2_mean': 0.0028586278203874826, 'reward_distinct_2_std': 0.008085420355200768, 'reward_judge_quality_mean': 0.10000000149011612, 'reward_judge_quality_std': 0.053452253341674805, 'reward_total_composite_mean': 0.016795290634036064, 'reward_total_composite_std': 0.02730741538107395, 'epoch': 0.04}

 29%|██▉       | 972/3300 [3:54:43<6:38:23, 10.27s/it]INFO 04-12 14:40:55 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 973/3300 [3:54:56<7:13:13, 11.17s/it]2026-04-12 14:41:08,789 | INFO | train_grpo_train | metrics_logged mode=train step=973

                                                      
{'loss': -0.0355, 'grad_norm': 2.1190357208251953, 'learning_rate': 7.054545454545455e-06, 'num_tokens': 1977325.0, 'completions/mean_length': 220.875, 'completions/min_length': 41.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 46.20000076293945, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.2139018177986145, 'rewards/meter/std': 0.3611465096473694, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.924515426158905, 'rewards/lexical_plausibility/std': 0.1314423382282257, 'rewards/judge_quality/mean': 0.41750001907348633, 'rewards/judge_quality/std': 0.3612182140350342, 'rewards/repeat_penalty/mean': 0.9901825189590454, 'rewards/repeat_penalty/std': 0.011737116612493992, 'rewards/near_duplicate_penalty/mean': 0.9906794428825378, 'rewards/near_duplicate_penalty/std': 0.011979267932474613, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9995018243789673, 'rewards/distinct_2/std': 0.0014091377379372716, 'rewards/total_composite/mean': 0.15802770853042603, 'rewards/total_composite/std': 0.31189027428627014, 'reward': 0.15802770853042603, 'reward_std': 0.31189030408859253, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08841715008020401, 'sampling/sampling_logp_difference/max': 1.142976999282837, 'sampling/importance_sampling_ratio/min': 0.31886833906173706, 'sampling/importance_sampling_ratio/mean': 0.992090106010437, 'sampling/importance_sampling_ratio/max': 1.6711084842681885, 'entropy': 0.2510576546192169, 'clip_ratio/low_mean': 0.035811505280435085, 'clip_ratio/low_min': 0.035811505280435085, 'clip_ratio/high_mean': 0.01884828880429268, 'clip_ratio/high_max': 0.01884828880429268, 'clip_ratio/region_mean': 0.054659794084727764, 'reward_total_mean': 0.15802770853042603, 'reward_meter_mean': 0.2139018177986145, 'reward_meter_std': 0.3611465096473694, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.924515426158905, 'reward_lexical_plausibility_std': 0.1314423382282257, 'reward_repeat_penalty_mean': 0.9901825189590454, 'reward_repeat_penalty_std': 0.011737116612493992, 'reward_near_duplicate_penalty_mean': 0.9906794428825378, 'reward_near_duplicate_penalty_std': 0.011979267932474613, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9995018243789673, 'reward_distinct_2_std': 0.0014091377379372716, 'reward_judge_quality_mean': 0.41750001907348633, 'reward_judge_quality_std': 0.3612182140350342, 'reward_total_composite_mean': 0.15802770853042603, 'reward_total_composite_std': 0.31189027428627014, 'epoch': 0.04}

 29%|██▉       | 973/3300 [3:54:56<7:13:13, 11.17s/it]INFO 04-12 14:41:09 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 974/3300 [3:55:05<6:41:31, 10.36s/it]2026-04-12 14:41:17,257 | INFO | train_grpo_train | metrics_logged mode=train step=974

                                                      
{'loss': 0.0214, 'grad_norm': 10.77816390991211, 'learning_rate': 7.0515151515151525e-06, 'num_tokens': 1979057.0, 'completions/mean_length': 55.5, 'completions/min_length': 50.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.5, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.5139241218566895, 'rewards/meter/std': 0.3450099527835846, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.8004724979400635, 'rewards/repeat_penalty/std': 0.051337871700525284, 'rewards/near_duplicate_penalty/mean': 0.8946834802627563, 'rewards/near_duplicate_penalty/std': 0.04043261706829071, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8946537375450134, 'rewards/distinct_2/std': 0.037854667752981186, 'rewards/total_composite/mean': 0.14144492149353027, 'rewards/total_composite/std': 0.11570578068494797, 'reward': 0.14144492149353027, 'reward_std': 0.11570577323436737, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07911837100982666, 'sampling/sampling_logp_difference/max': 2.6304514408111572, 'sampling/importance_sampling_ratio/min': 0.07204592972993851, 'sampling/importance_sampling_ratio/mean': 1.0089739561080933, 'sampling/importance_sampling_ratio/max': 1.8753541707992554, 'entropy': 0.4026899226009846, 'clip_ratio/low_mean': 0.05743443500250578, 'clip_ratio/low_min': 0.05743443500250578, 'clip_ratio/high_mean': 0.013762626331299543, 'clip_ratio/high_max': 0.013762626331299543, 'clip_ratio/region_mean': 0.07119706133380532, 'reward_total_mean': 0.14144492149353027, 'reward_meter_mean': 0.5139241218566895, 'reward_meter_std': 0.3450099527835846, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8004724979400635, 'reward_repeat_penalty_std': 0.051337871700525284, 'reward_near_duplicate_penalty_mean': 0.8946834802627563, 'reward_near_duplicate_penalty_std': 0.04043261706829071, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8946537375450134, 'reward_distinct_2_std': 0.037854667752981186, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.14144492149353027, 'reward_total_composite_std': 0.11570578068494797, 'epoch': 0.04}

 30%|██▉       | 974/3300 [3:55:05<6:41:31, 10.36s/it]INFO 04-12 14:41:17 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 975/3300 [3:55:13<6:14:27,  9.66s/it]2026-04-12 14:41:25,300 | INFO | train_grpo_train | metrics_logged mode=train step=975

                                                      
{'loss': 0.0847, 'grad_norm': 11.016331672668457, 'learning_rate': 7.048484848484849e-06, 'num_tokens': 1980982.0, 'completions/mean_length': 65.625, 'completions/min_length': 44.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.625, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.7261991500854492, 'rewards/meter/std': 0.3661386966705322, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.10690449178218842, 'rewards/repeat_penalty/mean': 0.8570008277893066, 'rewards/repeat_penalty/std': 0.05190880596637726, 'rewards/near_duplicate_penalty/mean': 0.9380278587341309, 'rewards/near_duplicate_penalty/std': 0.027473218739032745, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9130032658576965, 'rewards/distinct_2/std': 0.033487122505903244, 'rewards/total_composite/mean': 0.16555224359035492, 'rewards/total_composite/std': 0.10530142486095428, 'reward': 0.16555224359035492, 'reward_std': 0.10530142486095428, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12798751890659332, 'sampling/sampling_logp_difference/max': 2.6143364906311035, 'sampling/importance_sampling_ratio/min': 0.07321634888648987, 'sampling/importance_sampling_ratio/mean': 0.9907442927360535, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7357424721121788, 'clip_ratio/low_mean': 0.06466997228562832, 'clip_ratio/low_min': 0.06466997228562832, 'clip_ratio/high_mean': 0.04606099613010883, 'clip_ratio/high_max': 0.04606099613010883, 'clip_ratio/region_mean': 0.11073096841573715, 'reward_total_mean': 0.16555224359035492, 'reward_meter_mean': 0.7261991500854492, 'reward_meter_std': 0.3661386966705322, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8570008277893066, 'reward_repeat_penalty_std': 0.05190880596637726, 'reward_near_duplicate_penalty_mean': 0.9380278587341309, 'reward_near_duplicate_penalty_std': 0.027473218739032745, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9130032658576965, 'reward_distinct_2_std': 0.033487122505903244, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.10690449178218842, 'reward_total_composite_mean': 0.16555224359035492, 'reward_total_composite_std': 0.10530142486095428, 'epoch': 0.04}

 30%|██▉       | 975/3300 [3:55:13<6:14:27,  9.66s/it]INFO 04-12 14:41:25 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 976/3300 [3:55:23<6:23:08,  9.89s/it]2026-04-12 14:41:35,727 | INFO | train_grpo_train | metrics_logged mode=train step=976

                                                      
{'loss': -0.0092, 'grad_norm': 11.58521556854248, 'learning_rate': 7.045454545454546e-06, 'num_tokens': 1982764.0, 'completions/mean_length': 50.75, 'completions/min_length': 44.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.75, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.307443767786026, 'rewards/meter/std': 0.34788575768470764, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.9914896488189697, 'rewards/repeat_penalty/std': 0.013000753708183765, 'rewards/near_duplicate_penalty/mean': 0.9914896488189697, 'rewards/near_duplicate_penalty/std': 0.013000753708183765, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1296008825302124, 'rewards/total_composite/std': 0.15783850848674774, 'reward': 0.1296008825302124, 'reward_std': 0.15783850848674774, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09777948260307312, 'sampling/sampling_logp_difference/max': 1.714308738708496, 'sampling/importance_sampling_ratio/min': 0.18008817732334137, 'sampling/importance_sampling_ratio/mean': 1.0009520053863525, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.46424221247434616, 'clip_ratio/low_mean': 0.053921748884022236, 'clip_ratio/low_min': 0.053921748884022236, 'clip_ratio/high_mean': 0.042861031368374825, 'clip_ratio/high_max': 0.042861031368374825, 'clip_ratio/region_mean': 0.09678278025239706, 'reward_total_mean': 0.1296008825302124, 'reward_meter_mean': 0.307443767786026, 'reward_meter_std': 0.34788575768470764, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.9914896488189697, 'reward_repeat_penalty_std': 0.013000753708183765, 'reward_near_duplicate_penalty_mean': 0.9914896488189697, 'reward_near_duplicate_penalty_std': 0.013000753708183765, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.1296008825302124, 'reward_total_composite_std': 0.15783850848674774, 'epoch': 0.04}

 30%|██▉       | 976/3300 [3:55:23<6:23:08,  9.89s/it]INFO 04-12 14:41:35 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 977/3300 [3:55:32<6:06:33,  9.47s/it]2026-04-12 14:41:44,214 | INFO | train_grpo_train | metrics_logged mode=train step=977

                                                      
{'loss': 0.0323, 'grad_norm': 11.873543739318848, 'learning_rate': 7.0424242424242426e-06, 'num_tokens': 1984545.0, 'completions/mean_length': 60.625, 'completions/min_length': 55.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.625, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.5593838691711426, 'rewards/meter/std': 0.40533238649368286, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.510095477104187, 'rewards/repeat_penalty/std': 0.31443583965301514, 'rewards/near_duplicate_penalty/mean': 0.8125889301300049, 'rewards/near_duplicate_penalty/std': 0.14173707365989685, 'rewards/opening_diversity/mean': 0.828125, 'rewards/opening_diversity/std': 0.22097088396549225, 'rewards/distinct_2/mean': 0.7150894999504089, 'rewards/distinct_2/std': 0.22910983860492706, 'rewards/total_composite/mean': 0.17091941833496094, 'rewards/total_composite/std': 0.14195594191551208, 'reward': 0.17091941833496094, 'reward_std': 0.14195595681667328, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09142956882715225, 'sampling/sampling_logp_difference/max': 1.9169671535491943, 'sampling/importance_sampling_ratio/min': 0.14705228805541992, 'sampling/importance_sampling_ratio/mean': 1.0105063915252686, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4898384101688862, 'clip_ratio/low_mean': 0.059813302010297775, 'clip_ratio/low_min': 0.059813302010297775, 'clip_ratio/high_mean': 0.02513905195519328, 'clip_ratio/high_max': 0.02513905195519328, 'clip_ratio/region_mean': 0.08495235396549106, 'reward_total_mean': 0.17091941833496094, 'reward_meter_mean': 0.5593838691711426, 'reward_meter_std': 0.40533238649368286, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.510095477104187, 'reward_repeat_penalty_std': 0.31443583965301514, 'reward_near_duplicate_penalty_mean': 0.8125889301300049, 'reward_near_duplicate_penalty_std': 0.14173707365989685, 'reward_opening_diversity_mean': 0.828125, 'reward_opening_diversity_std': 0.22097088396549225, 'reward_distinct_2_mean': 0.7150894999504089, 'reward_distinct_2_std': 0.22910983860492706, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.17091941833496094, 'reward_total_composite_std': 0.14195594191551208, 'epoch': 0.04}

 30%|██▉       | 977/3300 [3:55:32<6:06:33,  9.47s/it]INFO 04-12 14:41:44 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 978/3300 [3:55:45<6:47:37, 10.53s/it]2026-04-12 14:41:57,219 | INFO | train_grpo_train | metrics_logged mode=train step=978

                                                      
{'loss': -0.0316, 'grad_norm': 4.624334812164307, 'learning_rate': 7.039393939393941e-06, 'num_tokens': 1986310.0, 'completions/mean_length': 125.625, 'completions/min_length': 62.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 70.42857360839844, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.3001912534236908, 'rewards/meter/std': 0.2824583351612091, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.2357022762298584, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9679378867149353, 'rewards/lexical_plausibility/std': 0.09068537503480911, 'rewards/judge_quality/mean': 0.20875000953674316, 'rewards/judge_quality/std': 0.12494999170303345, 'rewards/repeat_penalty/mean': 0.4374067187309265, 'rewards/repeat_penalty/std': 0.2799793779850006, 'rewards/near_duplicate_penalty/mean': 0.776577889919281, 'rewards/near_duplicate_penalty/std': 0.1323293000459671, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.24775780737400055, 'rewards/distinct_2/mean': 0.6423899531364441, 'rewards/distinct_2/std': 0.2421913743019104, 'rewards/total_composite/mean': 0.07037614285945892, 'rewards/total_composite/std': 0.10844790935516357, 'reward': 0.07037614285945892, 'reward_std': 0.10844790935516357, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09270121157169342, 'sampling/sampling_logp_difference/max': 2.46293568611145, 'sampling/importance_sampling_ratio/min': 0.0851845070719719, 'sampling/importance_sampling_ratio/mean': 0.992294430732727, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.33466818183660507, 'clip_ratio/low_mean': 0.028240297455340624, 'clip_ratio/low_min': 0.028240297455340624, 'clip_ratio/high_mean': 0.018362977541983128, 'clip_ratio/high_max': 0.018362977541983128, 'clip_ratio/region_mean': 0.04660327499732375, 'reward_total_mean': 0.07037614285945892, 'reward_meter_mean': 0.3001912534236908, 'reward_meter_std': 0.2824583351612091, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.2357022762298584, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9679378867149353, 'reward_lexical_plausibility_std': 0.09068537503480911, 'reward_repeat_penalty_mean': 0.4374067187309265, 'reward_repeat_penalty_std': 0.2799793779850006, 'reward_near_duplicate_penalty_mean': 0.776577889919281, 'reward_near_duplicate_penalty_std': 0.1323293000459671, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.24775780737400055, 'reward_distinct_2_mean': 0.6423899531364441, 'reward_distinct_2_std': 0.2421913743019104, 'reward_judge_quality_mean': 0.20875000953674316, 'reward_judge_quality_std': 0.12494999170303345, 'reward_total_composite_mean': 0.07037614285945892, 'reward_total_composite_std': 0.10844790935516357, 'epoch': 0.04}

 30%|██▉       | 978/3300 [3:55:45<6:47:37, 10.53s/it]INFO 04-12 14:41:57 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 979/3300 [3:55:58<7:18:51, 11.35s/it]2026-04-12 14:42:10,479 | INFO | train_grpo_train | metrics_logged mode=train step=979

                                                      
{'loss': -0.0143, 'grad_norm': 7.157508850097656, 'learning_rate': 7.036363636363637e-06, 'num_tokens': 1987942.0, 'completions/mean_length': 102.0, 'completions/min_length': 36.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 43.42857360839844, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.30481141805648804, 'rewards/meter/std': 0.3104307949542999, '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/lexical_plausibility/mean': 0.9627835750579834, 'rewards/lexical_plausibility/std': 0.08294543623924255, 'rewards/judge_quality/mean': 0.38999998569488525, 'rewards/judge_quality/std': 0.2684878408908844, 'rewards/repeat_penalty/mean': 0.6679938435554504, 'rewards/repeat_penalty/std': 0.2514175474643707, 'rewards/near_duplicate_penalty/mean': 0.8570210933685303, 'rewards/near_duplicate_penalty/std': 0.09549296647310257, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.8596557378768921, 'rewards/distinct_2/std': 0.1111566424369812, 'rewards/total_composite/mean': 0.15191224217414856, 'rewards/total_composite/std': 0.1981995850801468, 'reward': 0.15191224217414856, 'reward_std': 0.1981995850801468, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13293837010860443, 'sampling/sampling_logp_difference/max': 2.4101366996765137, 'sampling/importance_sampling_ratio/min': 0.08980302512645721, 'sampling/importance_sampling_ratio/mean': 1.008516788482666, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4877931773662567, 'clip_ratio/low_mean': 0.05374071467667818, 'clip_ratio/low_min': 0.05374071467667818, 'clip_ratio/high_mean': 0.035756501369178295, 'clip_ratio/high_max': 0.035756501369178295, 'clip_ratio/region_mean': 0.08949721604585648, 'reward_total_mean': 0.15191224217414856, 'reward_meter_mean': 0.30481141805648804, 'reward_meter_std': 0.3104307949542999, '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_lexical_plausibility_mean': 0.9627835750579834, 'reward_lexical_plausibility_std': 0.08294543623924255, 'reward_repeat_penalty_mean': 0.6679938435554504, 'reward_repeat_penalty_std': 0.2514175474643707, 'reward_near_duplicate_penalty_mean': 0.8570210933685303, 'reward_near_duplicate_penalty_std': 0.09549296647310257, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.8596557378768921, 'reward_distinct_2_std': 0.1111566424369812, 'reward_judge_quality_mean': 0.38999998569488525, 'reward_judge_quality_std': 0.2684878408908844, 'reward_total_composite_mean': 0.15191224217414856, 'reward_total_composite_std': 0.1981995850801468, 'epoch': 0.04}

 30%|██▉       | 979/3300 [3:55:58<7:18:51, 11.35s/it]INFO 04-12 14:42:10 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 980/3300 [3:56:07<6:57:39, 10.80s/it]2026-04-12 14:42:19,987 | INFO | train_grpo_train | metrics_logged mode=train step=980

                                                      
{'loss': -0.0519, 'grad_norm': 8.318925857543945, 'learning_rate': 7.033333333333334e-06, 'num_tokens': 1989996.0, 'completions/mean_length': 87.75, 'completions/min_length': 62.0, 'completions/max_length': 107.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 87.75, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 107.0, 'rewards/meter/mean': 0.6406745314598083, 'rewards/meter/std': 0.3038019835948944, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.17251639068126678, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/repeat_penalty/mean': 0.7011221647262573, 'rewards/repeat_penalty/std': 0.19219118356704712, 'rewards/near_duplicate_penalty/mean': 0.8614804148674011, 'rewards/near_duplicate_penalty/std': 0.0857333168387413, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8023130893707275, 'rewards/distinct_2/std': 0.15280494093894958, 'rewards/total_composite/mean': 0.17078441381454468, 'rewards/total_composite/std': 0.11170023679733276, 'reward': 0.17078441381454468, 'reward_std': 0.11170024424791336, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11083410680294037, 'sampling/sampling_logp_difference/max': 1.7277414798736572, 'sampling/importance_sampling_ratio/min': 0.17768526077270508, 'sampling/importance_sampling_ratio/mean': 0.9954089522361755, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.587139930576086, 'clip_ratio/low_mean': 0.05488824797794223, 'clip_ratio/low_min': 0.05488824797794223, 'clip_ratio/high_mean': 0.04168161191046238, 'clip_ratio/high_max': 0.04168161191046238, 'clip_ratio/region_mean': 0.09656985988840461, 'reward_total_mean': 0.17078441381454468, 'reward_meter_mean': 0.6406745314598083, 'reward_meter_std': 0.3038019835948944, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.17251639068126678, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7011221647262573, 'reward_repeat_penalty_std': 0.19219118356704712, 'reward_near_duplicate_penalty_mean': 0.8614804148674011, 'reward_near_duplicate_penalty_std': 0.0857333168387413, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8023130893707275, 'reward_distinct_2_std': 0.15280494093894958, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.17078441381454468, 'reward_total_composite_std': 0.11170023679733276, 'epoch': 0.04}

 30%|██▉       | 980/3300 [3:56:07<6:57:39, 10.80s/it]INFO 04-12 14:42:20 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 981/3300 [3:56:16<6:26:20, 10.00s/it]2026-04-12 14:42:28,107 | INFO | train_grpo_train | metrics_logged mode=train step=981

                                                      
{'loss': 0.0034, 'grad_norm': 12.545379638671875, 'learning_rate': 7.030303030303031e-06, 'num_tokens': 1991690.0, 'completions/mean_length': 53.75, 'completions/min_length': 42.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.75, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.5458482503890991, 'rewards/meter/std': 0.4694296717643738, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.10690449178218842, 'rewards/repeat_penalty/mean': 0.8491734266281128, 'rewards/repeat_penalty/std': 0.12976500391960144, 'rewards/near_duplicate_penalty/mean': 0.9190741777420044, 'rewards/near_duplicate_penalty/std': 0.04579726979136467, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9496719837188721, 'rewards/distinct_2/std': 0.06028076633810997, 'rewards/total_composite/mean': 0.13214111328125, 'rewards/total_composite/std': 0.1428379863500595, 'reward': 0.13214111328125, 'reward_std': 0.1428380012512207, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10618890821933746, 'sampling/sampling_logp_difference/max': 1.7462825775146484, 'sampling/importance_sampling_ratio/min': 0.17442113161087036, 'sampling/importance_sampling_ratio/mean': 1.0013306140899658, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7513331100344658, 'clip_ratio/low_mean': 0.061195218469947577, 'clip_ratio/low_min': 0.061195218469947577, 'clip_ratio/high_mean': 0.06303606275469065, 'clip_ratio/high_max': 0.06303606275469065, 'clip_ratio/region_mean': 0.12423128122463822, 'reward_total_mean': 0.13214111328125, 'reward_meter_mean': 0.5458482503890991, 'reward_meter_std': 0.4694296717643738, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8491734266281128, 'reward_repeat_penalty_std': 0.12976500391960144, 'reward_near_duplicate_penalty_mean': 0.9190741777420044, 'reward_near_duplicate_penalty_std': 0.04579726979136467, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9496719837188721, 'reward_distinct_2_std': 0.06028076633810997, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.10690449178218842, 'reward_total_composite_mean': 0.13214111328125, 'reward_total_composite_std': 0.1428379863500595, 'epoch': 0.04}

 30%|██▉       | 981/3300 [3:56:16<6:26:20, 10.00s/it]INFO 04-12 14:42:28 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 982/3300 [3:56:28<6:58:18, 10.83s/it]2026-04-12 14:42:40,874 | INFO | train_grpo_train | metrics_logged mode=train step=982

                                                      
{'loss': -0.0143, 'grad_norm': 3.270921230316162, 'learning_rate': 7.027272727272728e-06, 'num_tokens': 1993399.0, 'completions/mean_length': 96.625, 'completions/min_length': 30.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 37.28571701049805, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.26701539754867554, 'rewards/meter/std': 0.31732940673828125, '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/lexical_plausibility/mean': 0.9545800685882568, 'rewards/lexical_plausibility/std': 0.12846696376800537, 'rewards/judge_quality/mean': 0.5850000381469727, 'rewards/judge_quality/std': 0.377548485994339, 'rewards/repeat_penalty/mean': 0.9925603866577148, 'rewards/repeat_penalty/std': 0.014823234640061855, 'rewards/near_duplicate_penalty/mean': 0.9925603866577148, 'rewards/near_duplicate_penalty/std': 0.014823234640061855, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.17170283198356628, 'rewards/total_composite/std': 0.28969505429267883, 'reward': 0.17170283198356628, 'reward_std': 0.28969505429267883, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13185356557369232, 'sampling/sampling_logp_difference/max': 1.8705167770385742, 'sampling/importance_sampling_ratio/min': 0.32935336232185364, 'sampling/importance_sampling_ratio/mean': 1.0352216958999634, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5752882957458496, 'clip_ratio/low_mean': 0.08289665216580033, 'clip_ratio/low_min': 0.08289665216580033, 'clip_ratio/high_mean': 0.021557487081736326, 'clip_ratio/high_max': 0.021557487081736326, 'clip_ratio/region_mean': 0.10445413924753666, 'reward_total_mean': 0.17170283198356628, 'reward_meter_mean': 0.26701539754867554, 'reward_meter_std': 0.31732940673828125, '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_lexical_plausibility_mean': 0.9545800685882568, 'reward_lexical_plausibility_std': 0.12846696376800537, 'reward_repeat_penalty_mean': 0.9925603866577148, 'reward_repeat_penalty_std': 0.014823234640061855, 'reward_near_duplicate_penalty_mean': 0.9925603866577148, 'reward_near_duplicate_penalty_std': 0.014823234640061855, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5850000381469727, 'reward_judge_quality_std': 0.377548485994339, 'reward_total_composite_mean': 0.17170283198356628, 'reward_total_composite_std': 0.28969505429267883, 'epoch': 0.04}

 30%|██▉       | 982/3300 [3:56:28<6:58:18, 10.83s/it]INFO 04-12 14:42:41 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 983/3300 [3:56:36<6:25:51,  9.99s/it]2026-04-12 14:42:48,914 | INFO | train_grpo_train | metrics_logged mode=train step=983

                                                      
{'loss': 0.0233, 'grad_norm': 13.882776260375977, 'learning_rate': 7.024242424242424e-06, 'num_tokens': 1995099.0, 'completions/mean_length': 53.5, 'completions/min_length': 44.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.5, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.6158996820449829, 'rewards/meter/std': 0.39532360434532166, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4087499976158142, 'rewards/judge_quality/std': 0.24121345579624176, 'rewards/repeat_penalty/mean': 0.7647675275802612, 'rewards/repeat_penalty/std': 0.27327078580856323, 'rewards/near_duplicate_penalty/mean': 0.8891000747680664, 'rewards/near_duplicate_penalty/std': 0.11054614186286926, 'rewards/opening_diversity/mean': 0.921875, 'rewards/opening_diversity/std': 0.22097088396549225, 'rewards/distinct_2/mean': 0.8779939413070679, 'rewards/distinct_2/std': 0.14640764892101288, 'rewards/total_composite/mean': 0.2117781937122345, 'rewards/total_composite/std': 0.16315396130084991, 'reward': 0.2117781937122345, 'reward_std': 0.16315394639968872, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12032093107700348, 'sampling/sampling_logp_difference/max': 1.7155661582946777, 'sampling/importance_sampling_ratio/min': 0.1798618584871292, 'sampling/importance_sampling_ratio/mean': 1.0023165941238403, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5884692668914795, 'clip_ratio/low_mean': 0.06312001775950193, 'clip_ratio/low_min': 0.06312001775950193, 'clip_ratio/high_mean': 0.041774327866733074, 'clip_ratio/high_max': 0.041774327866733074, 'clip_ratio/region_mean': 0.10489434562623501, 'reward_total_mean': 0.2117781937122345, 'reward_meter_mean': 0.6158996820449829, 'reward_meter_std': 0.39532360434532166, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7647675275802612, 'reward_repeat_penalty_std': 0.27327078580856323, 'reward_near_duplicate_penalty_mean': 0.8891000747680664, 'reward_near_duplicate_penalty_std': 0.11054614186286926, 'reward_opening_diversity_mean': 0.921875, 'reward_opening_diversity_std': 0.22097088396549225, 'reward_distinct_2_mean': 0.8779939413070679, 'reward_distinct_2_std': 0.14640764892101288, 'reward_judge_quality_mean': 0.4087499976158142, 'reward_judge_quality_std': 0.24121345579624176, 'reward_total_composite_mean': 0.2117781937122345, 'reward_total_composite_std': 0.16315396130084991, 'epoch': 0.04}

 30%|██▉       | 983/3300 [3:56:36<6:25:51,  9.99s/it]INFO 04-12 14:42:49 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 984/3300 [3:56:44<5:54:02,  9.17s/it]2026-04-12 14:42:56,173 | INFO | train_grpo_train | metrics_logged mode=train step=984

                                                      
{'loss': 0.0716, 'grad_norm': 26.248443603515625, 'learning_rate': 7.021212121212122e-06, 'num_tokens': 1996609.0, 'completions/mean_length': 25.75, 'completions/min_length': 23.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.75, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.7959403991699219, 'rewards/meter/std': 0.3071911633014679, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.984375, 'rewards/lexical_plausibility/std': 0.0289318785071373, 'rewards/judge_quality/mean': 0.7275000214576721, 'rewards/judge_quality/std': 0.3564407229423523, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6245854496955872, 'rewards/total_composite/std': 0.360934853553772, 'reward': 0.6245854496955872, 'reward_std': 0.3609348237514496, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14976392686367035, 'sampling/sampling_logp_difference/max': 1.3163299560546875, 'sampling/importance_sampling_ratio/min': 0.2681175172328949, 'sampling/importance_sampling_ratio/mean': 1.0189403295516968, 'sampling/importance_sampling_ratio/max': 1.968184471130371, 'entropy': 0.8199795559048653, 'clip_ratio/low_mean': 0.041183135006576777, 'clip_ratio/low_min': 0.041183135006576777, 'clip_ratio/high_mean': 0.07003465062007308, 'clip_ratio/high_max': 0.07003465062007308, 'clip_ratio/region_mean': 0.11121778562664986, 'reward_total_mean': 0.6245854496955872, 'reward_meter_mean': 0.7959403991699219, 'reward_meter_std': 0.3071911633014679, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.984375, 'reward_lexical_plausibility_std': 0.0289318785071373, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7275000214576721, 'reward_judge_quality_std': 0.3564407229423523, 'reward_total_composite_mean': 0.6245854496955872, 'reward_total_composite_std': 0.360934853553772, 'epoch': 0.04}

 30%|██▉       | 984/3300 [3:56:44<5:54:02,  9.17s/it]INFO 04-12 14:42:56 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 985/3300 [3:56:57<6:46:38, 10.54s/it]2026-04-12 14:43:09,920 | INFO | train_grpo_train | metrics_logged mode=train step=985

                                                      
{'loss': -0.0885, 'grad_norm': 2.12365460395813, 'learning_rate': 7.018181818181818e-06, 'num_tokens': 1998284.0, 'completions/mean_length': 172.375, 'completions/min_length': 57.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 59.16666793823242, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.6781532168388367, 'rewards/meter/std': 0.3348863124847412, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9298042058944702, 'rewards/lexical_plausibility/std': 0.11668843030929565, 'rewards/judge_quality/mean': 0.20000000298023224, 'rewards/judge_quality/std': 0.13093073666095734, 'rewards/repeat_penalty/mean': 0.891778826713562, 'rewards/repeat_penalty/std': 0.11939821392297745, 'rewards/near_duplicate_penalty/mean': 0.9382267594337463, 'rewards/near_duplicate_penalty/std': 0.05382538586854935, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9469859600067139, 'rewards/distinct_2/std': 0.08032916486263275, 'rewards/total_composite/mean': 0.13072469830513, 'rewards/total_composite/std': 0.11394831538200378, 'reward': 0.13072469830513, 'reward_std': 0.11394830793142319, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1436917781829834, 'sampling/sampling_logp_difference/max': 1.8405656814575195, 'sampling/importance_sampling_ratio/min': 0.15872761607170105, 'sampling/importance_sampling_ratio/mean': 1.003473162651062, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5855006724596024, 'clip_ratio/low_mean': 0.02892102301120758, 'clip_ratio/low_min': 0.02892102301120758, 'clip_ratio/high_mean': 0.05759432818740606, 'clip_ratio/high_max': 0.05759432818740606, 'clip_ratio/region_mean': 0.08651535119861364, 'reward_total_mean': 0.13072469830513, 'reward_meter_mean': 0.6781532168388367, 'reward_meter_std': 0.3348863124847412, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9298042058944702, 'reward_lexical_plausibility_std': 0.11668843030929565, 'reward_repeat_penalty_mean': 0.891778826713562, 'reward_repeat_penalty_std': 0.11939821392297745, 'reward_near_duplicate_penalty_mean': 0.9382267594337463, 'reward_near_duplicate_penalty_std': 0.05382538586854935, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9469859600067139, 'reward_distinct_2_std': 0.08032916486263275, 'reward_judge_quality_mean': 0.20000000298023224, 'reward_judge_quality_std': 0.13093073666095734, 'reward_total_composite_mean': 0.13072469830513, 'reward_total_composite_std': 0.11394831538200378, 'epoch': 0.04}

 30%|██▉       | 985/3300 [3:56:57<6:46:38, 10.54s/it]INFO 04-12 14:43:10 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 986/3300 [3:57:11<7:25:48, 11.56s/it]2026-04-12 14:43:23,847 | INFO | train_grpo_train | metrics_logged mode=train step=986

                                                      
{'loss': -0.1223, 'grad_norm': 2.4662036895751953, 'learning_rate': 7.015151515151516e-06, 'num_tokens': 2000193.0, 'completions/mean_length': 185.625, 'completions/min_length': 74.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 76.83333587646484, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.7226679921150208, 'rewards/meter/std': 0.39732110500335693, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.2357022762298584, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9465062022209167, 'rewards/lexical_plausibility/std': 0.10204039514064789, 'rewards/judge_quality/mean': 0.39625000953674316, 'rewards/judge_quality/std': 0.3077075481414795, 'rewards/repeat_penalty/mean': 0.680860161781311, 'rewards/repeat_penalty/std': 0.2433183193206787, 'rewards/near_duplicate_penalty/mean': 0.8877695798873901, 'rewards/near_duplicate_penalty/std': 0.09313931316137314, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.18600596487522125, 'rewards/distinct_2/mean': 0.8297022581100464, 'rewards/distinct_2/std': 0.1726221740245819, 'rewards/total_composite/mean': 0.37500452995300293, 'rewards/total_composite/std': 0.3034655749797821, 'reward': 0.37500452995300293, 'reward_std': 0.3034655451774597, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1175990104675293, 'sampling/sampling_logp_difference/max': 1.858527660369873, 'sampling/importance_sampling_ratio/min': 0.15590199828147888, 'sampling/importance_sampling_ratio/mean': 0.9983144998550415, 'sampling/importance_sampling_ratio/max': 1.9845352172851562, 'entropy': 0.4068913646042347, 'clip_ratio/low_mean': 0.021170715801417828, 'clip_ratio/low_min': 0.021170715801417828, 'clip_ratio/high_mean': 0.05509928334504366, 'clip_ratio/high_max': 0.05509928334504366, 'clip_ratio/region_mean': 0.07626999914646149, 'reward_total_mean': 0.37500452995300293, 'reward_meter_mean': 0.7226679921150208, 'reward_meter_std': 0.39732110500335693, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.2357022762298584, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9465062022209167, 'reward_lexical_plausibility_std': 0.10204039514064789, 'reward_repeat_penalty_mean': 0.680860161781311, 'reward_repeat_penalty_std': 0.2433183193206787, 'reward_near_duplicate_penalty_mean': 0.8877695798873901, 'reward_near_duplicate_penalty_std': 0.09313931316137314, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.18600596487522125, 'reward_distinct_2_mean': 0.8297022581100464, 'reward_distinct_2_std': 0.1726221740245819, 'reward_judge_quality_mean': 0.39625000953674316, 'reward_judge_quality_std': 0.3077075481414795, 'reward_total_composite_mean': 0.37500452995300293, 'reward_total_composite_std': 0.3034655749797821, 'epoch': 0.04}

 30%|██▉       | 986/3300 [3:57:11<7:25:48, 11.56s/it]INFO 04-12 14:43:24 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 987/3300 [3:57:24<7:42:56, 12.01s/it]2026-04-12 14:43:36,907 | INFO | train_grpo_train | metrics_logged mode=train step=987

                                                      
{'loss': -0.0806, 'grad_norm': 2.222475290298462, 'learning_rate': 7.0121212121212126e-06, 'num_tokens': 2001640.0, 'completions/mean_length': 225.875, 'completions/min_length': 50.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 54.20000076293945, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.5086429119110107, 'rewards/meter/std': 0.422344833612442, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.902845025062561, 'rewards/lexical_plausibility/std': 0.13641053438186646, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.16690459847450256, 'rewards/repeat_penalty/mean': 0.7505574226379395, 'rewards/repeat_penalty/std': 0.15078061819076538, 'rewards/near_duplicate_penalty/mean': 0.9147303700447083, 'rewards/near_duplicate_penalty/std': 0.07815375179052353, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9087716341018677, 'rewards/distinct_2/std': 0.12953445315361023, 'rewards/total_composite/mean': 0.16314013302326202, 'rewards/total_composite/std': 0.1624373346567154, 'reward': 0.16314013302326202, 'reward_std': 0.1624373346567154, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10534392297267914, 'sampling/sampling_logp_difference/max': 0.9642497301101685, 'sampling/importance_sampling_ratio/min': 0.3812691569328308, 'sampling/importance_sampling_ratio/mean': 1.0275280475616455, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.424203522503376, 'clip_ratio/low_mean': 0.01666666753590107, 'clip_ratio/low_min': 0.01666666753590107, 'clip_ratio/high_mean': 0.03353535337373614, 'clip_ratio/high_max': 0.03353535337373614, 'clip_ratio/region_mean': 0.05020202090963721, 'reward_total_mean': 0.16314013302326202, 'reward_meter_mean': 0.5086429119110107, 'reward_meter_std': 0.422344833612442, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.902845025062561, 'reward_lexical_plausibility_std': 0.13641053438186646, 'reward_repeat_penalty_mean': 0.7505574226379395, 'reward_repeat_penalty_std': 0.15078061819076538, 'reward_near_duplicate_penalty_mean': 0.9147303700447083, 'reward_near_duplicate_penalty_std': 0.07815375179052353, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9087716341018677, 'reward_distinct_2_std': 0.12953445315361023, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.16690459847450256, 'reward_total_composite_mean': 0.16314013302326202, 'reward_total_composite_std': 0.1624373346567154, 'epoch': 0.04}

 30%|██▉       | 987/3300 [3:57:24<7:42:56, 12.01s/it]INFO 04-12 14:43:37 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 988/3300 [3:57:41<8:42:01, 13.55s/it]2026-04-12 14:43:54,044 | INFO | train_grpo_train | metrics_logged mode=train step=988

                                                      
{'loss': -0.0906, 'grad_norm': 3.833721160888672, 'learning_rate': 7.00909090909091e-06, 'num_tokens': 2003407.0, 'completions/mean_length': 185.875, 'completions/min_length': 73.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 77.16667175292969, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.6221957206726074, 'rewards/meter/std': 0.44464191794395447, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9379913210868835, 'rewards/lexical_plausibility/std': 0.1160426214337349, 'rewards/judge_quality/mean': 0.4424999952316284, 'rewards/judge_quality/std': 0.3340979516506195, 'rewards/repeat_penalty/mean': 0.8773944973945618, 'rewards/repeat_penalty/std': 0.12275677174329758, 'rewards/near_duplicate_penalty/mean': 0.9386776089668274, 'rewards/near_duplicate_penalty/std': 0.051305800676345825, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9614783525466919, 'rewards/distinct_2/std': 0.0421803779900074, 'rewards/total_composite/mean': 0.36288678646087646, 'rewards/total_composite/std': 0.37206077575683594, 'reward': 0.36288678646087646, 'reward_std': 0.37206074595451355, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1109168604016304, 'sampling/sampling_logp_difference/max': 1.5353846549987793, 'sampling/importance_sampling_ratio/min': 0.21537283062934875, 'sampling/importance_sampling_ratio/mean': 1.0210031270980835, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4782083071768284, 'clip_ratio/low_mean': 0.036940517369657755, 'clip_ratio/low_min': 0.036940517369657755, 'clip_ratio/high_mean': 0.04129467345774174, 'clip_ratio/high_max': 0.04129467345774174, 'clip_ratio/region_mean': 0.07823519082739949, 'reward_total_mean': 0.36288678646087646, 'reward_meter_mean': 0.6221957206726074, 'reward_meter_std': 0.44464191794395447, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9379913210868835, 'reward_lexical_plausibility_std': 0.1160426214337349, 'reward_repeat_penalty_mean': 0.8773944973945618, 'reward_repeat_penalty_std': 0.12275677174329758, 'reward_near_duplicate_penalty_mean': 0.9386776089668274, 'reward_near_duplicate_penalty_std': 0.051305800676345825, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9614783525466919, 'reward_distinct_2_std': 0.0421803779900074, 'reward_judge_quality_mean': 0.4424999952316284, 'reward_judge_quality_std': 0.3340979516506195, 'reward_total_composite_mean': 0.36288678646087646, 'reward_total_composite_std': 0.37206077575683594, 'epoch': 0.04}

 30%|██▉       | 988/3300 [3:57:42<8:42:01, 13.55s/it]INFO 04-12 14:43:54 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 989/3300 [3:57:55<8:44:21, 13.61s/it]2026-04-12 14:44:07,813 | INFO | train_grpo_train | metrics_logged mode=train step=989

                                                      
{'loss': -0.0647, 'grad_norm': 5.70667028427124, 'learning_rate': 7.006060606060606e-06, 'num_tokens': 2004873.0, 'completions/mean_length': 223.25, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 50.0, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.45600587129592896, 'rewards/meter/std': 0.4731103181838989, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/lexical_plausibility/mean': 0.9052602648735046, 'rewards/lexical_plausibility/std': 0.13334181904792786, 'rewards/judge_quality/mean': 0.23749999701976776, 'rewards/judge_quality/std': 0.18077215552330017, 'rewards/repeat_penalty/mean': 0.8469921350479126, 'rewards/repeat_penalty/std': 0.13217680156230927, 'rewards/near_duplicate_penalty/mean': 0.9327105283737183, 'rewards/near_duplicate_penalty/std': 0.05177021771669388, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9046897888183594, 'rewards/distinct_2/std': 0.10607904195785522, 'rewards/total_composite/mean': 0.12616285681724548, 'rewards/total_composite/std': 0.18038931488990784, 'reward': 0.12616285681724548, 'reward_std': 0.18038929998874664, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10798030346632004, 'sampling/sampling_logp_difference/max': 1.5350217819213867, 'sampling/importance_sampling_ratio/min': 0.21545101702213287, 'sampling/importance_sampling_ratio/mean': 1.004711389541626, 'sampling/importance_sampling_ratio/max': 1.9663946628570557, 'entropy': 0.32983867451548576, 'clip_ratio/low_mean': 0.02533244714140892, 'clip_ratio/low_min': 0.02533244714140892, 'clip_ratio/high_mean': 0.03373056510463357, 'clip_ratio/high_max': 0.03373056510463357, 'clip_ratio/region_mean': 0.05906301224604249, 'reward_total_mean': 0.12616285681724548, 'reward_meter_mean': 0.45600587129592896, 'reward_meter_std': 0.4731103181838989, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_lexical_plausibility_mean': 0.9052602648735046, 'reward_lexical_plausibility_std': 0.13334181904792786, 'reward_repeat_penalty_mean': 0.8469921350479126, 'reward_repeat_penalty_std': 0.13217680156230927, 'reward_near_duplicate_penalty_mean': 0.9327105283737183, 'reward_near_duplicate_penalty_std': 0.05177021771669388, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9046897888183594, 'reward_distinct_2_std': 0.10607904195785522, 'reward_judge_quality_mean': 0.23749999701976776, 'reward_judge_quality_std': 0.18077215552330017, 'reward_total_composite_mean': 0.12616285681724548, 'reward_total_composite_std': 0.18038931488990784, 'epoch': 0.04}

 30%|██▉       | 989/3300 [3:57:55<8:44:21, 13.61s/it]INFO 04-12 14:44:08 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 990/3300 [3:58:08<8:40:01, 13.51s/it]2026-04-12 14:44:21,071 | INFO | train_grpo_train | metrics_logged mode=train step=990

                                                      
{'loss': -0.041, 'grad_norm': 1.3340327739715576, 'learning_rate': 7.0030303030303035e-06, 'num_tokens': 2006295.0, 'completions/mean_length': 337.75, 'completions/min_length': 41.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.625, 'completions/mean_terminated_length': 47.333335876464844, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.3125881552696228, 'rewards/meter/std': 0.34360480308532715, '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/lexical_plausibility/mean': 0.826207160949707, 'rewards/lexical_plausibility/std': 0.1467028707265854, 'rewards/judge_quality/mean': 0.1875, 'rewards/judge_quality/std': 0.1922609806060791, 'rewards/repeat_penalty/mean': 0.928420901298523, 'rewards/repeat_penalty/std': 0.059653110802173615, 'rewards/near_duplicate_penalty/mean': 0.9394897222518921, 'rewards/near_duplicate_penalty/std': 0.04542161524295807, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9878213405609131, 'rewards/distinct_2/std': 0.02975465916097164, 'rewards/total_composite/mean': 0.07774575054645538, 'rewards/total_composite/std': 0.1380847990512848, 'reward': 0.07774575054645538, 'reward_std': 0.1380847841501236, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15079817175865173, 'sampling/sampling_logp_difference/max': 1.9642105102539062, 'sampling/importance_sampling_ratio/min': 0.14026658236980438, 'sampling/importance_sampling_ratio/mean': 0.9765504598617554, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2385721057653427, 'clip_ratio/low_mean': 0.006818181835114956, 'clip_ratio/low_min': 0.006818181835114956, 'clip_ratio/high_mean': 0.05799310840666294, 'clip_ratio/high_max': 0.05799310840666294, 'clip_ratio/region_mean': 0.0648112902417779, 'reward_total_mean': 0.07774575054645538, 'reward_meter_mean': 0.3125881552696228, 'reward_meter_std': 0.34360480308532715, '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_lexical_plausibility_mean': 0.826207160949707, 'reward_lexical_plausibility_std': 0.1467028707265854, 'reward_repeat_penalty_mean': 0.928420901298523, 'reward_repeat_penalty_std': 0.059653110802173615, 'reward_near_duplicate_penalty_mean': 0.9394897222518921, 'reward_near_duplicate_penalty_std': 0.04542161524295807, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9878213405609131, 'reward_distinct_2_std': 0.02975465916097164, 'reward_judge_quality_mean': 0.1875, 'reward_judge_quality_std': 0.1922609806060791, 'reward_total_composite_mean': 0.07774575054645538, 'reward_total_composite_std': 0.1380847990512848, 'epoch': 0.04}

 30%|███       | 990/3300 [3:58:09<8:40:01, 13.51s/it]INFO 04-12 14:44:21 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 991/3300 [3:58:22<8:43:12, 13.60s/it]2026-04-12 14:44:34,867 | INFO | train_grpo_train | metrics_logged mode=train step=991

                                                      
{'loss': -0.2059, 'grad_norm': 1.2185224294662476, 'learning_rate': 7e-06, 'num_tokens': 2011194.0, 'completions/mean_length': 467.375, 'completions/min_length': 400.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 461.0000305175781, 'completions/min_terminated_length': 400.0, 'completions/max_terminated_length': 503.0, 'rewards/meter/mean': 0.7228045463562012, 'rewards/meter/std': 0.3669188320636749, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0971859022974968, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.125, 'rewards/judge_quality/std': 0.046291008591651917, 'rewards/repeat_penalty/mean': 0.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 0.2874869406223297, 'rewards/near_duplicate_penalty/std': 0.16659756004810333, 'rewards/opening_diversity/mean': 0.46581366658210754, 'rewards/opening_diversity/std': 0.3762345314025879, 'rewards/distinct_2/mean': 0.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.0455997996032238, 'rewards/total_composite/std': 0.025717701762914658, 'reward': 0.0455997996032238, 'reward_std': 0.02571769990026951, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.020445222035050392, 'sampling/sampling_logp_difference/max': 4.094364643096924, 'sampling/importance_sampling_ratio/min': 0.016666332259774208, 'sampling/importance_sampling_ratio/mean': 1.001975417137146, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.09206820605322719, 'clip_ratio/low_mean': 0.00502943410538137, 'clip_ratio/low_min': 0.00502943410538137, 'clip_ratio/high_mean': 0.008704663021489978, 'clip_ratio/high_max': 0.008704663021489978, 'clip_ratio/region_mean': 0.013734097126871347, 'reward_total_mean': 0.0455997996032238, 'reward_meter_mean': 0.7228045463562012, 'reward_meter_std': 0.3669188320636749, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0971859022974968, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.0, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 0.2874869406223297, 'reward_near_duplicate_penalty_std': 0.16659756004810333, 'reward_opening_diversity_mean': 0.46581366658210754, 'reward_opening_diversity_std': 0.3762345314025879, 'reward_distinct_2_mean': 0.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.125, 'reward_judge_quality_std': 0.046291008591651917, 'reward_total_composite_mean': 0.0455997996032238, 'reward_total_composite_std': 0.025717701762914658, 'epoch': 0.04}

 30%|███       | 991/3300 [3:58:22<8:43:12, 13.60s/it]INFO 04-12 14:44:35 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 992/3300 [3:58:35<8:38:06, 13.47s/it]2026-04-12 14:44:48,043 | INFO | train_grpo_train | metrics_logged mode=train step=992

                                                      
{'loss': 0.0252, 'grad_norm': 2.940246343612671, 'learning_rate': 6.996969696969698e-06, 'num_tokens': 2016326.0, 'completions/mean_length': 422.5, 'completions/min_length': 390.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 422.5, 'completions/min_terminated_length': 390.0, 'completions/max_terminated_length': 512.0, 'rewards/meter/mean': 0.8226900696754456, 'rewards/meter/std': 0.19365568459033966, 'rewards/count_adherence/mean': 0.5795454978942871, 'rewards/count_adherence/std': 0.06763851642608643, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.13750000298023224, 'rewards/judge_quality/std': 0.0353553406894207, 'rewards/repeat_penalty/mean': 0.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 0.5123528838157654, 'rewards/near_duplicate_penalty/std': 0.1367569863796234, 'rewards/opening_diversity/mean': 0.77724289894104, 'rewards/opening_diversity/std': 0.02403104119002819, 'rewards/distinct_2/mean': 0.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.06489282846450806, 'rewards/total_composite/std': 0.024038584902882576, 'reward': 0.06489282846450806, 'reward_std': 0.024038584902882576, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.020665442571043968, 'sampling/sampling_logp_difference/max': 1.7199337482452393, 'sampling/importance_sampling_ratio/min': 0.18318268656730652, 'sampling/importance_sampling_ratio/mean': 1.0008816719055176, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.10681440634652972, 'clip_ratio/low_mean': 0.009256523801013827, 'clip_ratio/low_min': 0.009256523801013827, 'clip_ratio/high_mean': 0.012108491035178304, 'clip_ratio/high_max': 0.012108491035178304, 'clip_ratio/region_mean': 0.02136501483619213, 'reward_total_mean': 0.06489282846450806, 'reward_meter_mean': 0.8226900696754456, 'reward_meter_std': 0.19365568459033966, 'reward_count_adherence_mean': 0.5795454978942871, 'reward_count_adherence_std': 0.06763851642608643, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.0, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 0.5123528838157654, 'reward_near_duplicate_penalty_std': 0.1367569863796234, 'reward_opening_diversity_mean': 0.77724289894104, 'reward_opening_diversity_std': 0.02403104119002819, 'reward_distinct_2_mean': 0.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.13750000298023224, 'reward_judge_quality_std': 0.0353553406894207, 'reward_total_composite_mean': 0.06489282846450806, 'reward_total_composite_std': 0.024038584902882576, 'epoch': 0.04}

 30%|███       | 992/3300 [3:58:36<8:38:06, 13.47s/it]INFO 04-12 14:44:48 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 993/3300 [3:58:49<8:44:12, 13.63s/it]2026-04-12 14:45:02,064 | INFO | train_grpo_train | metrics_logged mode=train step=993

                                                      
{'loss': -0.0614, 'grad_norm': 2.8004815578460693, 'learning_rate': 6.993939393939394e-06, 'num_tokens': 2017871.0, 'completions/mean_length': 223.125, 'completions/min_length': 46.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 49.79999923706055, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.08863702416419983, 'rewards/meter/std': 0.10776276886463165, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9058598875999451, 'rewards/lexical_plausibility/std': 0.1252550631761551, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.20830951631069183, 'rewards/repeat_penalty/mean': 0.9608862996101379, 'rewards/repeat_penalty/std': 0.03716663271188736, 'rewards/near_duplicate_penalty/mean': 0.9719347357749939, 'rewards/near_duplicate_penalty/std': 0.034501202404499054, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9888646602630615, 'rewards/distinct_2/std': 0.026741670444607735, 'rewards/total_composite/mean': 0.03434469923377037, 'rewards/total_composite/std': 0.051052313297986984, 'reward': 0.03434469923377037, 'reward_std': 0.05105230584740639, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13200277090072632, 'sampling/sampling_logp_difference/max': 1.546842098236084, 'sampling/importance_sampling_ratio/min': 0.3204887807369232, 'sampling/importance_sampling_ratio/mean': 1.0063825845718384, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.46162882447242737, 'clip_ratio/low_mean': 0.024109620600938797, 'clip_ratio/low_min': 0.024109620600938797, 'clip_ratio/high_mean': 0.04244799492880702, 'clip_ratio/high_max': 0.04244799492880702, 'clip_ratio/region_mean': 0.06655761552974582, 'reward_total_mean': 0.03434469923377037, 'reward_meter_mean': 0.08863702416419983, 'reward_meter_std': 0.10776276886463165, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9058598875999451, 'reward_lexical_plausibility_std': 0.1252550631761551, 'reward_repeat_penalty_mean': 0.9608862996101379, 'reward_repeat_penalty_std': 0.03716663271188736, 'reward_near_duplicate_penalty_mean': 0.9719347357749939, 'reward_near_duplicate_penalty_std': 0.034501202404499054, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9888646602630615, 'reward_distinct_2_std': 0.026741670444607735, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.20830951631069183, 'reward_total_composite_mean': 0.03434469923377037, 'reward_total_composite_std': 0.051052313297986984, 'epoch': 0.04}

 30%|███       | 993/3300 [3:58:50<8:44:12, 13.63s/it]INFO 04-12 14:45:02 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 994/3300 [3:59:04<8:54:01, 13.89s/it]2026-04-12 14:45:16,570 | INFO | train_grpo_train | metrics_logged mode=train step=994

                                                      
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 6.990909090909092e-06, 'num_tokens': 2019583.0, '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': 0.7204117774963379, 'rewards/meter/std': 0.30648282170295715, 'rewards/count_adherence/mean': 0.8602941036224365, 'rewards/count_adherence/std': 0.10863599181175232, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.995192289352417, 'rewards/lexical_plausibility/std': 0.013598216697573662, 'rewards/judge_quality/mean': 0.13750000298023224, 'rewards/judge_quality/std': 0.0353553406894207, 'rewards/repeat_penalty/mean': 0.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 0.30068665742874146, 'rewards/near_duplicate_penalty/std': 0.16507866978645325, 'rewards/opening_diversity/mean': 0.29391634464263916, 'rewards/opening_diversity/std': 0.3674202263355255, 'rewards/distinct_2/mean': 0.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.08356542885303497, 'rewards/total_composite/std': 0.045433610677719116, 'reward': 0.08356542885303497, 'reward_std': 0.04543360695242882, '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': 0.08356542885303497, 'reward_meter_mean': 0.7204117774963379, 'reward_meter_std': 0.30648282170295715, 'reward_count_adherence_mean': 0.8602941036224365, 'reward_count_adherence_std': 0.10863599181175232, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.995192289352417, 'reward_lexical_plausibility_std': 0.013598216697573662, 'reward_repeat_penalty_mean': 0.0, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 0.30068665742874146, 'reward_near_duplicate_penalty_std': 0.16507866978645325, 'reward_opening_diversity_mean': 0.29391634464263916, 'reward_opening_diversity_std': 0.3674202263355255, 'reward_distinct_2_mean': 0.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.13750000298023224, 'reward_judge_quality_std': 0.0353553406894207, 'reward_total_composite_mean': 0.08356542885303497, 'reward_total_composite_std': 0.045433610677719116, 'epoch': 0.04}

 30%|███       | 994/3300 [3:59:04<8:54:01, 13.89s/it]INFO 04-12 14:45:16 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 995/3300 [3:59:18<8:51:25, 13.83s/it]2026-04-12 14:45:30,263 | INFO | train_grpo_train | metrics_logged mode=train step=995

                                                      
{'loss': -0.0717, 'grad_norm': 4.690229415893555, 'learning_rate': 6.987878787878788e-06, 'num_tokens': 2021367.0, 'completions/mean_length': 112.0, 'completions/min_length': 48.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 54.857147216796875, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.64249587059021, 'rewards/meter/std': 0.37544283270835876, '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/lexical_plausibility/mean': 0.9674509763717651, 'rewards/lexical_plausibility/std': 0.09206251800060272, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.12464233487844467, 'rewards/repeat_penalty/mean': 0.23540544509887695, 'rewards/repeat_penalty/std': 0.06962842494249344, 'rewards/near_duplicate_penalty/mean': 0.822376012802124, 'rewards/near_duplicate_penalty/std': 0.08210719376802444, 'rewards/opening_diversity/mean': 0.375, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.7789473533630371, 'rewards/distinct_2/std': 0.11217062175273895, 'rewards/total_composite/mean': 0.24081265926361084, 'rewards/total_composite/std': 0.15572069585323334, 'reward': 0.24081265926361084, 'reward_std': 0.15572069585323334, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11030396074056625, 'sampling/sampling_logp_difference/max': 4.377264022827148, 'sampling/importance_sampling_ratio/min': 0.012559674680233002, 'sampling/importance_sampling_ratio/mean': 1.0015673637390137, 'sampling/importance_sampling_ratio/max': 1.9557359218597412, 'entropy': 0.41459524631500244, 'clip_ratio/low_mean': 0.015321316663175821, 'clip_ratio/low_min': 0.015321316663175821, 'clip_ratio/high_mean': 0.06096663139760494, 'clip_ratio/high_max': 0.06096663139760494, 'clip_ratio/region_mean': 0.07628794806078076, 'reward_total_mean': 0.24081265926361084, 'reward_meter_mean': 0.64249587059021, 'reward_meter_std': 0.37544283270835876, '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_lexical_plausibility_mean': 0.9674509763717651, 'reward_lexical_plausibility_std': 0.09206251800060272, 'reward_repeat_penalty_mean': 0.23540544509887695, 'reward_repeat_penalty_std': 0.06962842494249344, 'reward_near_duplicate_penalty_mean': 0.822376012802124, 'reward_near_duplicate_penalty_std': 0.08210719376802444, 'reward_opening_diversity_mean': 0.375, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.7789473533630371, 'reward_distinct_2_std': 0.11217062175273895, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.12464233487844467, 'reward_total_composite_mean': 0.24081265926361084, 'reward_total_composite_std': 0.15572069585323334, 'epoch': 0.04}

 30%|███       | 995/3300 [3:59:18<8:51:25, 13.83s/it]INFO 04-12 14:45:30 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 996/3300 [3:59:31<8:43:11, 13.62s/it]2026-04-12 14:45:43,403 | INFO | train_grpo_train | metrics_logged mode=train step=996

                                                      
{'loss': -0.0734, 'grad_norm': 5.3192524909973145, 'learning_rate': 6.984848484848485e-06, 'num_tokens': 2023094.0, 'completions/mean_length': 115.875, 'completions/min_length': 54.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 59.28571701049805, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.6954101324081421, 'rewards/meter/std': 0.4064273238182068, '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/lexical_plausibility/mean': 0.9541934728622437, 'rewards/lexical_plausibility/std': 0.1295604705810547, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.12464234977960587, 'rewards/repeat_penalty/mean': 0.2111881971359253, 'rewards/repeat_penalty/std': 0.08781961351633072, 'rewards/near_duplicate_penalty/mean': 0.7852634787559509, 'rewards/near_duplicate_penalty/std': 0.11110790073871613, 'rewards/opening_diversity/mean': 0.375, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.7237294912338257, 'rewards/distinct_2/std': 0.1848858743906021, 'rewards/total_composite/mean': 0.1936955749988556, 'rewards/total_composite/std': 0.13327328860759735, 'reward': 0.1936955749988556, 'reward_std': 0.13327328860759735, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10754402726888657, 'sampling/sampling_logp_difference/max': 1.513790488243103, 'sampling/importance_sampling_ratio/min': 0.22007420659065247, 'sampling/importance_sampling_ratio/mean': 1.0193158388137817, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5054878517985344, 'clip_ratio/low_mean': 0.02899590227752924, 'clip_ratio/low_min': 0.02899590227752924, 'clip_ratio/high_mean': 0.034066562075167894, 'clip_ratio/high_max': 0.034066562075167894, 'clip_ratio/region_mean': 0.06306246435269713, 'reward_total_mean': 0.1936955749988556, 'reward_meter_mean': 0.6954101324081421, 'reward_meter_std': 0.4064273238182068, '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_lexical_plausibility_mean': 0.9541934728622437, 'reward_lexical_plausibility_std': 0.1295604705810547, 'reward_repeat_penalty_mean': 0.2111881971359253, 'reward_repeat_penalty_std': 0.08781961351633072, 'reward_near_duplicate_penalty_mean': 0.7852634787559509, 'reward_near_duplicate_penalty_std': 0.11110790073871613, 'reward_opening_diversity_mean': 0.375, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.7237294912338257, 'reward_distinct_2_std': 0.1848858743906021, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.12464234977960587, 'reward_total_composite_mean': 0.1936955749988556, 'reward_total_composite_std': 0.13327328860759735, 'epoch': 0.04}

 30%|███       | 996/3300 [3:59:31<8:43:11, 13.62s/it]INFO 04-12 14:45:43 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 997/3300 [3:59:44<8:39:54, 13.55s/it]2026-04-12 14:45:56,766 | INFO | train_grpo_train | metrics_logged mode=train step=997

                                                      
{'loss': -0.0604, 'grad_norm': 5.503066539764404, 'learning_rate': 6.981818181818183e-06, 'num_tokens': 2024833.0, 'completions/mean_length': 116.375, 'completions/min_length': 43.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 59.857147216796875, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.3735466003417969, 'rewards/meter/std': 0.37652871012687683, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9714733362197876, 'rewards/lexical_plausibility/std': 0.05957445502281189, 'rewards/judge_quality/mean': 0.4312499761581421, 'rewards/judge_quality/std': 0.2034303992986679, 'rewards/repeat_penalty/mean': 0.9736570119857788, 'rewards/repeat_penalty/std': 0.02958020195364952, 'rewards/near_duplicate_penalty/mean': 0.9806067943572998, 'rewards/near_duplicate_penalty/std': 0.020286794751882553, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9928774833679199, 'rewards/distinct_2/std': 0.02014549821615219, 'rewards/total_composite/mean': 0.1606295257806778, 'rewards/total_composite/std': 0.1760026067495346, 'reward': 0.1606295257806778, 'reward_std': 0.1760026067495346, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13005565106868744, 'sampling/sampling_logp_difference/max': 2.1428146362304688, 'sampling/importance_sampling_ratio/min': 0.11732415854930878, 'sampling/importance_sampling_ratio/mean': 0.9885691404342651, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.35430509597063065, 'clip_ratio/low_mean': 0.05167744169011712, 'clip_ratio/low_min': 0.05167744169011712, 'clip_ratio/high_mean': 0.01801715325564146, 'clip_ratio/high_max': 0.01801715325564146, 'clip_ratio/region_mean': 0.06969459494575858, 'reward_total_mean': 0.1606295257806778, 'reward_meter_mean': 0.3735466003417969, 'reward_meter_std': 0.37652871012687683, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9714733362197876, 'reward_lexical_plausibility_std': 0.05957445502281189, 'reward_repeat_penalty_mean': 0.9736570119857788, 'reward_repeat_penalty_std': 0.02958020195364952, 'reward_near_duplicate_penalty_mean': 0.9806067943572998, 'reward_near_duplicate_penalty_std': 0.020286794751882553, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9928774833679199, 'reward_distinct_2_std': 0.02014549821615219, 'reward_judge_quality_mean': 0.4312499761581421, 'reward_judge_quality_std': 0.2034303992986679, 'reward_total_composite_mean': 0.1606295257806778, 'reward_total_composite_std': 0.1760026067495346, 'epoch': 0.04}

 30%|███       | 997/3300 [3:59:44<8:39:54, 13.55s/it]INFO 04-12 14:45:57 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 998/3300 [3:59:57<8:33:45, 13.39s/it]2026-04-12 14:46:09,788 | INFO | train_grpo_train | metrics_logged mode=train step=998

                                                      
{'loss': -0.0778, 'grad_norm': 3.5188589096069336, 'learning_rate': 6.978787878787879e-06, 'num_tokens': 2026240.0, 'completions/mean_length': 89.875, 'completions/min_length': 27.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 29.571430206298828, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.9511470794677734, 'rewards/meter/std': 0.05642685666680336, '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/lexical_plausibility/mean': 0.9348700046539307, 'rewards/lexical_plausibility/std': 0.1206575408577919, 'rewards/judge_quality/mean': 0.752500057220459, 'rewards/judge_quality/std': 0.32805708050727844, 'rewards/repeat_penalty/mean': 0.6575287580490112, 'rewards/repeat_penalty/std': 0.19856330752372742, 'rewards/near_duplicate_penalty/mean': 0.9058570861816406, 'rewards/near_duplicate_penalty/std': 0.11396578699350357, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.90974360704422, 'rewards/distinct_2/std': 0.12743444740772247, 'rewards/total_composite/mean': 0.7146866321563721, 'rewards/total_composite/std': 0.329754114151001, 'reward': 0.7146866321563721, 'reward_std': 0.3297540843486786, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11943802982568741, 'sampling/sampling_logp_difference/max': 2.5625662803649902, 'sampling/importance_sampling_ratio/min': 0.07710660994052887, 'sampling/importance_sampling_ratio/mean': 0.9861419200897217, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3236630205065012, 'clip_ratio/low_mean': 0.012096773833036423, 'clip_ratio/low_min': 0.012096773833036423, 'clip_ratio/high_mean': 0.08018164988607168, 'clip_ratio/high_max': 0.08018164988607168, 'clip_ratio/region_mean': 0.0922784237191081, 'reward_total_mean': 0.7146866321563721, 'reward_meter_mean': 0.9511470794677734, 'reward_meter_std': 0.05642685666680336, '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_lexical_plausibility_mean': 0.9348700046539307, 'reward_lexical_plausibility_std': 0.1206575408577919, 'reward_repeat_penalty_mean': 0.6575287580490112, 'reward_repeat_penalty_std': 0.19856330752372742, 'reward_near_duplicate_penalty_mean': 0.9058570861816406, 'reward_near_duplicate_penalty_std': 0.11396578699350357, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.90974360704422, 'reward_distinct_2_std': 0.12743444740772247, 'reward_judge_quality_mean': 0.752500057220459, 'reward_judge_quality_std': 0.32805708050727844, 'reward_total_composite_mean': 0.7146866321563721, 'reward_total_composite_std': 0.329754114151001, 'epoch': 0.04}

 30%|███       | 998/3300 [3:59:57<8:33:45, 13.39s/it]INFO 04-12 14:46:10 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 999/3300 [4:00:11<8:37:42, 13.50s/it]2026-04-12 14:46:23,659 | INFO | train_grpo_train | metrics_logged mode=train step=999

                                                      
{'loss': -0.118, 'grad_norm': 2.2460145950317383, 'learning_rate': 6.975757575757577e-06, 'num_tokens': 2027889.0, 'completions/mean_length': 233.125, 'completions/min_length': 45.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 65.80000305175781, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 90.0, 'rewards/meter/mean': 0.3810475468635559, 'rewards/meter/std': 0.34137123823165894, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.17251639068126678, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.8971705436706543, 'rewards/lexical_plausibility/std': 0.14915716648101807, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.22320714592933655, 'rewards/repeat_penalty/mean': 0.9307105541229248, 'rewards/repeat_penalty/std': 0.10673901438713074, 'rewards/near_duplicate_penalty/mean': 0.9803764224052429, 'rewards/near_duplicate_penalty/std': 0.025375651195645332, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9785828590393066, 'rewards/distinct_2/std': 0.03611748293042183, 'rewards/total_composite/mean': 0.11651533097028732, 'rewards/total_composite/std': 0.13248947262763977, 'reward': 0.11651533097028732, 'reward_std': 0.13248947262763977, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1157555878162384, 'sampling/sampling_logp_difference/max': 1.4340004920959473, 'sampling/importance_sampling_ratio/min': 0.24378632009029388, 'sampling/importance_sampling_ratio/mean': 1.005470633506775, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4111524745821953, 'clip_ratio/low_mean': 0.023634454235434532, 'clip_ratio/low_min': 0.023634454235434532, 'clip_ratio/high_mean': 0.02964559430256486, 'clip_ratio/high_max': 0.02964559430256486, 'clip_ratio/region_mean': 0.05328004853799939, 'reward_total_mean': 0.11651533097028732, 'reward_meter_mean': 0.3810475468635559, 'reward_meter_std': 0.34137123823165894, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.17251639068126678, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.8971705436706543, 'reward_lexical_plausibility_std': 0.14915716648101807, 'reward_repeat_penalty_mean': 0.9307105541229248, 'reward_repeat_penalty_std': 0.10673901438713074, 'reward_near_duplicate_penalty_mean': 0.9803764224052429, 'reward_near_duplicate_penalty_std': 0.025375651195645332, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9785828590393066, 'reward_distinct_2_std': 0.03611748293042183, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.22320714592933655, 'reward_total_composite_mean': 0.11651533097028732, 'reward_total_composite_std': 0.13248947262763977, 'epoch': 0.04}

 30%|███       | 999/3300 [4:00:11<8:37:42, 13.50s/it]INFO 04-12 14:46:23 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 1000/3300 [4:00:21<7:55:20, 12.40s/it]2026-04-12 14:46:33,380 | INFO | train_grpo_train | metrics_logged mode=train step=1000

                                                       
{'loss': 0.0321, 'grad_norm': 8.04367733001709, 'learning_rate': 6.9727272727272735e-06, 'num_tokens': 2029980.0, 'completions/mean_length': 90.375, 'completions/min_length': 86.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 90.375, 'completions/min_terminated_length': 86.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.4259508550167084, 'rewards/meter/std': 0.33860698342323303, '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/lexical_plausibility/mean': 0.9866071343421936, 'rewards/lexical_plausibility/std': 0.03788072615861893, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/repeat_penalty/mean': 0.770172119140625, 'rewards/repeat_penalty/std': 0.24134013056755066, 'rewards/near_duplicate_penalty/mean': 0.9159718751907349, 'rewards/near_duplicate_penalty/std': 0.06320098042488098, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.8654276728630066, 'rewards/distinct_2/std': 0.1894504278898239, 'rewards/total_composite/mean': 0.14330320060253143, 'rewards/total_composite/std': 0.15179672837257385, 'reward': 0.14330320060253143, 'reward_std': 0.15179671347141266, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11694381386041641, 'sampling/sampling_logp_difference/max': 1.7405447959899902, 'sampling/importance_sampling_ratio/min': 0.17542479932308197, 'sampling/importance_sampling_ratio/mean': 1.0054761171340942, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7396683618426323, 'clip_ratio/low_mean': 0.05706383567303419, 'clip_ratio/low_min': 0.05706383567303419, 'clip_ratio/high_mean': 0.06889655813574791, 'clip_ratio/high_max': 0.06889655813574791, 'clip_ratio/region_mean': 0.1259603938087821, 'reward_total_mean': 0.14330320060253143, 'reward_meter_mean': 0.4259508550167084, 'reward_meter_std': 0.33860698342323303, '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_lexical_plausibility_mean': 0.9866071343421936, 'reward_lexical_plausibility_std': 0.03788072615861893, 'reward_repeat_penalty_mean': 0.770172119140625, 'reward_repeat_penalty_std': 0.24134013056755066, 'reward_near_duplicate_penalty_mean': 0.9159718751907349, 'reward_near_duplicate_penalty_std': 0.06320098042488098, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.8654276728630066, 'reward_distinct_2_std': 0.1894504278898239, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.14330320060253143, 'reward_total_composite_std': 0.15179672837257385, 'epoch': 0.04}

 30%|███       | 1000/3300 [4:00:21<7:55:20, 12.40s/it]INFO 04-12 14:46:33 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


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

 15%|█▌        | 2/13 [00:11<01:02,  5.65s/it]

 23%|██▎       | 3/13 [00:22<01:18,  7.83s/it]

 31%|███       | 4/13 [00:33<01:21,  9.03s/it]

 38%|███▊      | 5/13 [00:44<01:18,  9.79s/it]

 46%|████▌     | 6/13 [00:55<01:12, 10.30s/it]

 54%|█████▍    | 7/13 [01:06<01:02, 10.44s/it]

 62%|██████▏   | 8/13 [01:19<00:56, 11.32s/it]

 69%|██████▉   | 9/13 [01:30<00:45, 11.27s/it]

 77%|███████▋  | 10/13 [01:41<00:33, 11.09s/it]

 85%|████████▍ | 11/13 [01:53<00:22, 11.20s/it]

 92%|█████████▏| 12/13 [02:04<00:11, 11.26s/it]

100%|██████████| 13/13 [02:15<00:00, 11.20s/it]2026-04-12 14:49:01,161 | INFO | train_grpo_train | metrics_logged mode=eval step=1000

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 147.7034, 'eval_samples_per_second': 0.704, 'eval_steps_per_second': 0.088, 'eval_num_tokens': 2029980.0, 'eval_completions/mean_length': 256.96153846153845, 'eval_completions/min_length': 44.07692307692308, 'eval_completions/max_length': 512.0, 'eval_completions/clipped_ratio': 0.27884615384615385, 'eval_completions/mean_terminated_length': 157.76264190673828, 'eval_completions/min_terminated_length': 44.07692307692308, 'eval_completions/max_terminated_length': 308.3076923076923, 'eval_rewards/meter/mean': 0.4314915342972829, 'eval_rewards/meter/std': 0.36755329828995925, 'eval_rewards/count_adherence/mean': 0.7969282819674566, 'eval_rewards/count_adherence/std': 0.2405730955875837, 'eval_rewards/arabic_clean/mean': 0.9615384615384616, 'eval_rewards/arabic_clean/std': 0.06700789240690377, 'eval_rewards/lexical_plausibility/mean': 0.981945853966933, 'eval_rewards/lexical_plausibility/std': 0.03274063712272506, 'eval_rewards/judge_quality/mean': 0.23701923053998214, 'eval_rewards/judge_quality/std': 0.18446642962785867, 'eval_rewards/repeat_penalty/mean': 0.3449514794808168, 'eval_rewards/repeat_penalty/std': 0.3808590930241805, 'eval_rewards/near_duplicate_penalty/mean': 0.6952358713516822, 'eval_rewards/near_duplicate_penalty/std': 0.24378586560487747, 'eval_rewards/opening_diversity/mean': 0.7192492668445294, 'eval_rewards/opening_diversity/std': 0.3173264849644441, 'eval_rewards/distinct_2/mean': 0.4128593011544301, 'eval_rewards/distinct_2/std': 0.41273009318571824, 'eval_rewards/total_composite/mean': 0.09094488534789819, 'eval_rewards/total_composite/std': 0.1357112294779374, 'eval_reward': 0.09094488534789819, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.03131374329901659, 'eval_sampling/sampling_logp_difference/max': 0.8352567232572116, 'eval_sampling/importance_sampling_ratio/min': 0.45060978027490467, 'eval_sampling/importance_sampling_ratio/mean': 1.0072243763850286, 'eval_sampling/importance_sampling_ratio/max': 1.5050187569398146, 'eval_entropy': 0.32988868539149946, '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': 0.09094488534789819, 'eval_reward_meter_mean': 0.4314915342972829, 'eval_reward_meter_std': 0.36755329828995925, 'eval_reward_count_adherence_mean': 0.7969282819674566, 'eval_reward_count_adherence_std': 0.2405730955875837, 'eval_reward_arabic_clean_mean': 0.9615384615384616, 'eval_reward_arabic_clean_std': 0.06700789240690377, 'eval_reward_lexical_plausibility_mean': 0.981945853966933, 'eval_reward_lexical_plausibility_std': 0.03274063712272506, 'eval_reward_repeat_penalty_mean': 0.3449514794808168, 'eval_reward_repeat_penalty_std': 0.3808590930241805, 'eval_reward_near_duplicate_penalty_mean': 0.6952358713516822, 'eval_reward_near_duplicate_penalty_std': 0.24378586560487747, 'eval_reward_opening_diversity_mean': 0.7192492668445294, 'eval_reward_opening_diversity_std': 0.3173264849644441, 'eval_reward_distinct_2_mean': 0.4128593011544301, 'eval_reward_distinct_2_std': 0.41273009318571824, 'eval_reward_judge_quality_mean': 0.23701923053998214, 'eval_reward_judge_quality_std': 0.18446642962785867, 'eval_reward_total_composite_mean': 0.09094488534789819, 'eval_reward_total_composite_std': 0.1357112294779374, 'epoch': 0.04}

 30%|███       | 1000/3300 [4:02:49<7:55:20, 12.40s/it]

100%|██████████| 13/13 [02:15<00:00, 11.20s/it]

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

 30%|███       | 1001/3300 [4:03:04<36:53:03, 57.76s/it]2026-04-12 14:49:16,966 | INFO | train_grpo_train | metrics_logged mode=train step=1001

                                                        
{'loss': -0.0375, 'grad_norm': 3.9714503288269043, 'learning_rate': 6.969696969696971e-06, 'num_tokens': 2031511.0, 'completions/mean_length': 105.375, 'completions/min_length': 43.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 47.28571701049805, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.28398823738098145, 'rewards/meter/std': 0.2592170834541321, '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/lexical_plausibility/mean': 0.970693826675415, 'rewards/lexical_plausibility/std': 0.08289031684398651, 'rewards/judge_quality/mean': 0.5049999952316284, 'rewards/judge_quality/std': 0.28972893953323364, 'rewards/repeat_penalty/mean': 0.9407724142074585, 'rewards/repeat_penalty/std': 0.10241445153951645, 'rewards/near_duplicate_penalty/mean': 0.9902974963188171, 'rewards/near_duplicate_penalty/std': 0.018226707354187965, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9807692170143127, 'rewards/distinct_2/std': 0.054392825812101364, 'rewards/total_composite/mean': 0.1736394762992859, 'rewards/total_composite/std': 0.19033192098140717, 'reward': 0.1736394762992859, 'reward_std': 0.19033192098140717, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13567040860652924, 'sampling/sampling_logp_difference/max': 1.5328576564788818, 'sampling/importance_sampling_ratio/min': 0.21591776609420776, 'sampling/importance_sampling_ratio/mean': 0.9989410042762756, 'sampling/importance_sampling_ratio/max': 1.9708611965179443, 'entropy': 0.6063445881009102, 'clip_ratio/low_mean': 0.08616565633565187, 'clip_ratio/low_min': 0.08616565633565187, 'clip_ratio/high_mean': 0.06323696300387383, 'clip_ratio/high_max': 0.06323696300387383, 'clip_ratio/region_mean': 0.1494026193395257, 'reward_total_mean': 0.1736394762992859, 'reward_meter_mean': 0.28398823738098145, 'reward_meter_std': 0.2592170834541321, '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_lexical_plausibility_mean': 0.970693826675415, 'reward_lexical_plausibility_std': 0.08289031684398651, 'reward_repeat_penalty_mean': 0.9407724142074585, 'reward_repeat_penalty_std': 0.10241445153951645, 'reward_near_duplicate_penalty_mean': 0.9902974963188171, 'reward_near_duplicate_penalty_std': 0.018226707354187965, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9807692170143127, 'reward_distinct_2_std': 0.054392825812101364, 'reward_judge_quality_mean': 0.5049999952316284, 'reward_judge_quality_std': 0.28972893953323364, 'reward_total_composite_mean': 0.1736394762992859, 'reward_total_composite_std': 0.19033192098140717, 'epoch': 0.04}

 30%|███       | 1001/3300 [4:03:04<36:53:03, 57.76s/it]INFO 04-12 14:49:17 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 1002/3300 [4:03:18<28:20:10, 44.39s/it]2026-04-12 14:49:30,174 | INFO | train_grpo_train | metrics_logged mode=train step=1002

                                                        
{'loss': -0.0211, 'grad_norm': 6.3390793800354, 'learning_rate': 6.966666666666667e-06, 'num_tokens': 2032904.0, 'completions/mean_length': 98.125, 'completions/min_length': 35.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 39.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.18136833608150482, 'rewards/meter/std': 0.3299355208873749, '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/lexical_plausibility/mean': 0.9596225023269653, 'rewards/lexical_plausibility/std': 0.11420483142137527, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.16035674512386322, 'rewards/repeat_penalty/mean': 0.7080590724945068, 'rewards/repeat_penalty/std': 0.12625540792942047, 'rewards/near_duplicate_penalty/mean': 0.8970104455947876, 'rewards/near_duplicate_penalty/std': 0.06268128752708435, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.8180110454559326, 'rewards/distinct_2/std': 0.13785549998283386, 'rewards/total_composite/mean': 0.041882578283548355, 'rewards/total_composite/std': 0.04958256706595421, 'reward': 0.041882578283548355, 'reward_std': 0.04958256706595421, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12469225376844406, 'sampling/sampling_logp_difference/max': 2.209979772567749, 'sampling/importance_sampling_ratio/min': 0.10970286279916763, 'sampling/importance_sampling_ratio/mean': 0.9950018525123596, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5036148615181446, 'clip_ratio/low_mean': 0.04411446349695325, 'clip_ratio/low_min': 0.04411446349695325, 'clip_ratio/high_mean': 0.048820249270647764, 'clip_ratio/high_max': 0.048820249270647764, 'clip_ratio/region_mean': 0.09293471276760101, 'reward_total_mean': 0.041882578283548355, 'reward_meter_mean': 0.18136833608150482, 'reward_meter_std': 0.3299355208873749, '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_lexical_plausibility_mean': 0.9596225023269653, 'reward_lexical_plausibility_std': 0.11420483142137527, 'reward_repeat_penalty_mean': 0.7080590724945068, 'reward_repeat_penalty_std': 0.12625540792942047, 'reward_near_duplicate_penalty_mean': 0.8970104455947876, 'reward_near_duplicate_penalty_std': 0.06268128752708435, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.8180110454559326, 'reward_distinct_2_std': 0.13785549998283386, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.16035674512386322, 'reward_total_composite_mean': 0.041882578283548355, 'reward_total_composite_std': 0.04958256706595421, 'epoch': 0.04}

 30%|███       | 1002/3300 [4:03:18<28:20:10, 44.39s/it]INFO 04-12 14:49:30 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 1003/3300 [4:03:31<22:20:58, 35.03s/it]2026-04-12 14:49:43,347 | INFO | train_grpo_train | metrics_logged mode=train step=1003

                                                        
{'loss': -0.1572, 'grad_norm': 3.547816276550293, 'learning_rate': 6.963636363636364e-06, 'num_tokens': 2035103.0, 'completions/mean_length': 155.875, 'completions/min_length': 85.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 105.00000762939453, 'completions/min_terminated_length': 85.0, 'completions/max_terminated_length': 120.0, 'rewards/meter/mean': 0.7965437173843384, 'rewards/meter/std': 0.30849045515060425, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9801627397537231, 'rewards/lexical_plausibility/std': 0.056108247488737106, 'rewards/judge_quality/mean': 0.17499999701976776, 'rewards/judge_quality/std': 0.11649647355079651, 'rewards/repeat_penalty/mean': 0.5346405506134033, 'rewards/repeat_penalty/std': 0.22748194634914398, 'rewards/near_duplicate_penalty/mean': 0.813302755355835, 'rewards/near_duplicate_penalty/std': 0.1031440794467926, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.6712074279785156, 'rewards/distinct_2/std': 0.21098695695400238, 'rewards/total_composite/mean': 0.10118478536605835, 'rewards/total_composite/std': 0.05983646959066391, 'reward': 0.10118478536605835, 'reward_std': 0.05983646959066391, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09340889751911163, 'sampling/sampling_logp_difference/max': 1.6200447082519531, 'sampling/importance_sampling_ratio/min': 0.19788984954357147, 'sampling/importance_sampling_ratio/mean': 1.0012426376342773, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3998628482222557, 'clip_ratio/low_mean': 0.012519201263785362, 'clip_ratio/low_min': 0.012519201263785362, 'clip_ratio/high_mean': 0.056228159461170435, 'clip_ratio/high_max': 0.056228159461170435, 'clip_ratio/region_mean': 0.0687473607249558, 'reward_total_mean': 0.10118478536605835, 'reward_meter_mean': 0.7965437173843384, 'reward_meter_std': 0.30849045515060425, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9801627397537231, 'reward_lexical_plausibility_std': 0.056108247488737106, 'reward_repeat_penalty_mean': 0.5346405506134033, 'reward_repeat_penalty_std': 0.22748194634914398, 'reward_near_duplicate_penalty_mean': 0.813302755355835, 'reward_near_duplicate_penalty_std': 0.1031440794467926, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.6712074279785156, 'reward_distinct_2_std': 0.21098695695400238, 'reward_judge_quality_mean': 0.17499999701976776, 'reward_judge_quality_std': 0.11649647355079651, 'reward_total_composite_mean': 0.10118478536605835, 'reward_total_composite_std': 0.05983646959066391, 'epoch': 0.04}

 30%|███       | 1003/3300 [4:03:31<22:20:58, 35.03s/it]INFO 04-12 14:49:43 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 1004/3300 [4:03:45<18:17:00, 28.67s/it]2026-04-12 14:49:57,180 | INFO | train_grpo_train | metrics_logged mode=train step=1004

                                                        
{'loss': -0.0675, 'grad_norm': 2.0703108310699463, 'learning_rate': 6.960606060606061e-06, 'num_tokens': 2036618.0, 'completions/mean_length': 228.375, 'completions/min_length': 49.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 58.20000076293945, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.34577107429504395, 'rewards/meter/std': 0.40819546580314636, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.8809272050857544, 'rewards/lexical_plausibility/std': 0.18054750561714172, 'rewards/judge_quality/mean': 0.20624999701976776, 'rewards/judge_quality/std': 0.1590990275144577, 'rewards/repeat_penalty/mean': 0.7737113237380981, 'rewards/repeat_penalty/std': 0.23971515893936157, 'rewards/near_duplicate_penalty/mean': 0.8543292284011841, 'rewards/near_duplicate_penalty/std': 0.12977567315101624, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9022855162620544, 'rewards/distinct_2/std': 0.13790582120418549, 'rewards/total_composite/mean': 0.062010399997234344, 'rewards/total_composite/std': 0.07427313923835754, 'reward': 0.062010399997234344, 'reward_std': 0.07427313923835754, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10309043526649475, 'sampling/sampling_logp_difference/max': 3.445627450942993, 'sampling/importance_sampling_ratio/min': 0.03188474848866463, 'sampling/importance_sampling_ratio/mean': 1.0156586170196533, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2901155650615692, 'clip_ratio/low_mean': 0.016960651613771915, 'clip_ratio/low_min': 0.016960651613771915, 'clip_ratio/high_mean': 0.030373553279787302, 'clip_ratio/high_max': 0.030373553279787302, 'clip_ratio/region_mean': 0.04733420489355922, 'reward_total_mean': 0.062010399997234344, 'reward_meter_mean': 0.34577107429504395, 'reward_meter_std': 0.40819546580314636, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.8809272050857544, 'reward_lexical_plausibility_std': 0.18054750561714172, 'reward_repeat_penalty_mean': 0.7737113237380981, 'reward_repeat_penalty_std': 0.23971515893936157, 'reward_near_duplicate_penalty_mean': 0.8543292284011841, 'reward_near_duplicate_penalty_std': 0.12977567315101624, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9022855162620544, 'reward_distinct_2_std': 0.13790582120418549, 'reward_judge_quality_mean': 0.20624999701976776, 'reward_judge_quality_std': 0.1590990275144577, 'reward_total_composite_mean': 0.062010399997234344, 'reward_total_composite_std': 0.07427313923835754, 'epoch': 0.04}

 30%|███       | 1004/3300 [4:03:45<18:17:00, 28.67s/it]INFO 04-12 14:49:57 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 1005/3300 [4:03:54<14:39:23, 22.99s/it]2026-04-12 14:50:06,926 | INFO | train_grpo_train | metrics_logged mode=train step=1005

                                                        
{'loss': -0.022, 'grad_norm': 6.288625240325928, 'learning_rate': 6.957575757575759e-06, 'num_tokens': 2039289.0, 'completions/mean_length': 132.875, 'completions/min_length': 119.0, 'completions/max_length': 148.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 132.875, 'completions/min_terminated_length': 119.0, 'completions/max_terminated_length': 148.0, 'rewards/meter/mean': 0.5801243185997009, 'rewards/meter/std': 0.43735790252685547, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.20000000298023224, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.12380699813365936, 'rewards/repeat_penalty/std': 0.12617744505405426, 'rewards/near_duplicate_penalty/mean': 0.6660910844802856, 'rewards/near_duplicate_penalty/std': 0.14987289905548096, 'rewards/opening_diversity/mean': 0.8791666030883789, 'rewards/opening_diversity/std': 0.09708039462566376, 'rewards/distinct_2/mean': 0.21273569762706757, 'rewards/distinct_2/std': 0.19339588284492493, 'rewards/total_composite/mean': 0.12219199538230896, 'rewards/total_composite/std': 0.11519946157932281, 'reward': 0.12219199538230896, 'reward_std': 0.11519946157932281, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.067280612885952, 'sampling/sampling_logp_difference/max': 1.8679523468017578, 'sampling/importance_sampling_ratio/min': 0.15443958342075348, 'sampling/importance_sampling_ratio/mean': 1.0054123401641846, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.33262037485837936, 'clip_ratio/low_mean': 0.03251151228323579, 'clip_ratio/low_min': 0.03251151228323579, 'clip_ratio/high_mean': 0.03645447944290936, 'clip_ratio/high_max': 0.03645447944290936, 'clip_ratio/region_mean': 0.06896599172614515, 'reward_total_mean': 0.12219199538230896, 'reward_meter_mean': 0.5801243185997009, 'reward_meter_std': 0.43735790252685547, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.09258200973272324, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.12380699813365936, 'reward_repeat_penalty_std': 0.12617744505405426, 'reward_near_duplicate_penalty_mean': 0.6660910844802856, 'reward_near_duplicate_penalty_std': 0.14987289905548096, 'reward_opening_diversity_mean': 0.8791666030883789, 'reward_opening_diversity_std': 0.09708039462566376, 'reward_distinct_2_mean': 0.21273569762706757, 'reward_distinct_2_std': 0.19339588284492493, 'reward_judge_quality_mean': 0.20000000298023224, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.12219199538230896, 'reward_total_composite_std': 0.11519946157932281, 'epoch': 0.04}

 30%|███       | 1005/3300 [4:03:54<14:39:23, 22.99s/it]INFO 04-12 14:50:07 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 1006/3300 [4:04:07<12:42:28, 19.94s/it]2026-04-12 14:50:19,748 | INFO | train_grpo_train | metrics_logged mode=train step=1006

                                                        
{'loss': -0.0034, 'grad_norm': 4.988742351531982, 'learning_rate': 6.954545454545455e-06, 'num_tokens': 2041138.0, 'completions/mean_length': 139.125, 'completions/min_length': 53.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 85.85714721679688, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 111.0, 'rewards/meter/mean': 0.47990572452545166, 'rewards/meter/std': 0.45544543862342834, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9599024057388306, 'rewards/lexical_plausibility/std': 0.11341319233179092, 'rewards/judge_quality/mean': 0.1875, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/repeat_penalty/mean': 0.5238727331161499, 'rewards/repeat_penalty/std': 0.35384947061538696, 'rewards/near_duplicate_penalty/mean': 0.79697585105896, 'rewards/near_duplicate_penalty/std': 0.20894178748130798, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.29124119877815247, 'rewards/distinct_2/mean': 0.6515576243400574, 'rewards/distinct_2/std': 0.3078213334083557, 'rewards/total_composite/mean': 0.08327089250087738, 'rewards/total_composite/std': 0.08128806948661804, 'reward': 0.08327089250087738, 'reward_std': 0.08128806203603745, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08831564337015152, 'sampling/sampling_logp_difference/max': 1.5496015548706055, 'sampling/importance_sampling_ratio/min': 0.2123325616121292, 'sampling/importance_sampling_ratio/mean': 1.0111733675003052, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4837537780404091, 'clip_ratio/low_mean': 0.025016087107360363, 'clip_ratio/low_min': 0.025016087107360363, 'clip_ratio/high_mean': 0.07756210002116859, 'clip_ratio/high_max': 0.07756210002116859, 'clip_ratio/region_mean': 0.10257818712852895, 'reward_total_mean': 0.08327089250087738, 'reward_meter_mean': 0.47990572452545166, 'reward_meter_std': 0.45544543862342834, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9599024057388306, 'reward_lexical_plausibility_std': 0.11341319233179092, 'reward_repeat_penalty_mean': 0.5238727331161499, 'reward_repeat_penalty_std': 0.35384947061538696, 'reward_near_duplicate_penalty_mean': 0.79697585105896, 'reward_near_duplicate_penalty_std': 0.20894178748130798, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.29124119877815247, 'reward_distinct_2_mean': 0.6515576243400574, 'reward_distinct_2_std': 0.3078213334083557, 'reward_judge_quality_mean': 0.1875, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.08327089250087738, 'reward_total_composite_std': 0.08128806948661804, 'epoch': 0.04}

 30%|███       | 1006/3300 [4:04:07<12:42:28, 19.94s/it]INFO 04-12 14:50:19 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1007/3300 [4:04:21<11:29:55, 18.05s/it]2026-04-12 14:50:33,392 | INFO | train_grpo_train | metrics_logged mode=train step=1007

                                                        
{'loss': -0.0664, 'grad_norm': 1.8367737531661987, 'learning_rate': 6.951515151515153e-06, 'num_tokens': 2042633.0, 'completions/mean_length': 281.875, 'completions/min_length': 49.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 51.75, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.7451859712600708, 'rewards/meter/std': 0.3604169487953186, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.375, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.850554347038269, 'rewards/lexical_plausibility/std': 0.16505010426044464, 'rewards/judge_quality/mean': 0.20874999463558197, 'rewards/judge_quality/std': 0.18098440766334534, 'rewards/repeat_penalty/mean': 0.9535529613494873, 'rewards/repeat_penalty/std': 0.09540250897407532, 'rewards/near_duplicate_penalty/mean': 0.9853535890579224, 'rewards/near_duplicate_penalty/std': 0.023111911490559578, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9983469247817993, 'rewards/distinct_2/std': 0.0046756030060350895, 'rewards/total_composite/mean': 0.11863739043474197, 'rewards/total_composite/std': 0.16427050530910492, 'reward': 0.11863739043474197, 'reward_std': 0.16427049040794373, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11549889296293259, 'sampling/sampling_logp_difference/max': 1.4535958766937256, 'sampling/importance_sampling_ratio/min': 0.2337283194065094, 'sampling/importance_sampling_ratio/mean': 1.0086487531661987, 'sampling/importance_sampling_ratio/max': 1.9590433835983276, 'entropy': 0.2860126793384552, 'clip_ratio/low_mean': 0.010204081423580647, 'clip_ratio/low_min': 0.010204081423580647, 'clip_ratio/high_mean': 0.043494854122400284, 'clip_ratio/high_max': 0.043494854122400284, 'clip_ratio/region_mean': 0.05369893554598093, 'reward_total_mean': 0.11863739043474197, 'reward_meter_mean': 0.7451859712600708, 'reward_meter_std': 0.3604169487953186, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.375, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.850554347038269, 'reward_lexical_plausibility_std': 0.16505010426044464, 'reward_repeat_penalty_mean': 0.9535529613494873, 'reward_repeat_penalty_std': 0.09540250897407532, 'reward_near_duplicate_penalty_mean': 0.9853535890579224, 'reward_near_duplicate_penalty_std': 0.023111911490559578, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9983469247817993, 'reward_distinct_2_std': 0.0046756030060350895, 'reward_judge_quality_mean': 0.20874999463558197, 'reward_judge_quality_std': 0.18098440766334534, 'reward_total_composite_mean': 0.11863739043474197, 'reward_total_composite_std': 0.16427050530910492, 'epoch': 0.04}

 31%|███       | 1007/3300 [4:04:21<11:29:55, 18.05s/it]INFO 04-12 14:50:33 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1008/3300 [4:04:35<10:40:42, 16.77s/it]2026-04-12 14:50:47,190 | INFO | train_grpo_train | metrics_logged mode=train step=1008

                                                        
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 6.948484848484849e-06, 'num_tokens': 2044201.0, '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': 0.7862903475761414, 'rewards/meter/std': 0.20163999497890472, 'rewards/count_adherence/mean': 0.4553571343421936, 'rewards/count_adherence/std': 0.16170331835746765, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9191277027130127, 'rewards/lexical_plausibility/std': 0.15767855942249298, 'rewards/judge_quality/mean': 0.125, 'rewards/judge_quality/std': 0.046291008591651917, 'rewards/repeat_penalty/mean': 0.02640357054769993, 'rewards/repeat_penalty/std': 0.06918352097272873, 'rewards/near_duplicate_penalty/mean': 0.4383179545402527, 'rewards/near_duplicate_penalty/std': 0.20757566392421722, 'rewards/opening_diversity/mean': 0.38697630167007446, 'rewards/opening_diversity/std': 0.3496139645576477, 'rewards/distinct_2/mean': 0.19237789511680603, 'rewards/distinct_2/std': 0.3637513518333435, 'rewards/total_composite/mean': 0.04245631396770477, 'rewards/total_composite/std': 0.02771882899105549, 'reward': 0.04245631396770477, 'reward_std': 0.02771882899105549, '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': 0.04245631396770477, 'reward_meter_mean': 0.7862903475761414, 'reward_meter_std': 0.20163999497890472, 'reward_count_adherence_mean': 0.4553571343421936, 'reward_count_adherence_std': 0.16170331835746765, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9191277027130127, 'reward_lexical_plausibility_std': 0.15767855942249298, 'reward_repeat_penalty_mean': 0.02640357054769993, 'reward_repeat_penalty_std': 0.06918352097272873, 'reward_near_duplicate_penalty_mean': 0.4383179545402527, 'reward_near_duplicate_penalty_std': 0.20757566392421722, 'reward_opening_diversity_mean': 0.38697630167007446, 'reward_opening_diversity_std': 0.3496139645576477, 'reward_distinct_2_mean': 0.19237789511680603, 'reward_distinct_2_std': 0.3637513518333435, 'reward_judge_quality_mean': 0.125, 'reward_judge_quality_std': 0.046291008591651917, 'reward_total_composite_mean': 0.04245631396770477, 'reward_total_composite_std': 0.02771882899105549, 'epoch': 0.04}

 31%|███       | 1008/3300 [4:04:35<10:40:42, 16.77s/it]INFO 04-12 14:50:47 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1009/3300 [4:04:48<10:05:18, 15.85s/it]2026-04-12 14:51:00,890 | INFO | train_grpo_train | metrics_logged mode=train step=1009

                                                        
{'loss': -0.0533, 'grad_norm': 6.875668048858643, 'learning_rate': 6.945454545454546e-06, 'num_tokens': 2046171.0, 'completions/mean_length': 136.25, 'completions/min_length': 78.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 82.5714340209961, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.5190771222114563, 'rewards/meter/std': 0.41183415055274963, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.2357022762298584, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/lexical_plausibility/mean': 0.9531154632568359, 'rewards/lexical_plausibility/std': 0.13260941207408905, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.12464234977960587, 'rewards/repeat_penalty/mean': 0.7970947027206421, 'rewards/repeat_penalty/std': 0.13408467173576355, 'rewards/near_duplicate_penalty/mean': 0.92569899559021, 'rewards/near_duplicate_penalty/std': 0.0491853728890419, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8782718181610107, 'rewards/distinct_2/std': 0.08077096194028854, 'rewards/total_composite/mean': 0.11063134670257568, 'rewards/total_composite/std': 0.12931987643241882, 'reward': 0.11063134670257568, 'reward_std': 0.12931987643241882, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11841348558664322, 'sampling/sampling_logp_difference/max': 1.5015144348144531, 'sampling/importance_sampling_ratio/min': 0.22279250621795654, 'sampling/importance_sampling_ratio/mean': 1.0072684288024902, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6490747630596161, 'clip_ratio/low_mean': 0.062433586455881596, 'clip_ratio/low_min': 0.062433586455881596, 'clip_ratio/high_mean': 0.04343796148896217, 'clip_ratio/high_max': 0.04343796148896217, 'clip_ratio/region_mean': 0.10587154794484377, 'reward_total_mean': 0.11063134670257568, 'reward_meter_mean': 0.5190771222114563, 'reward_meter_std': 0.41183415055274963, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.2357022762298584, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_lexical_plausibility_mean': 0.9531154632568359, 'reward_lexical_plausibility_std': 0.13260941207408905, 'reward_repeat_penalty_mean': 0.7970947027206421, 'reward_repeat_penalty_std': 0.13408467173576355, 'reward_near_duplicate_penalty_mean': 0.92569899559021, 'reward_near_duplicate_penalty_std': 0.0491853728890419, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8782718181610107, 'reward_distinct_2_std': 0.08077096194028854, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.12464234977960587, 'reward_total_composite_mean': 0.11063134670257568, 'reward_total_composite_std': 0.12931987643241882, 'epoch': 0.04}

 31%|███       | 1009/3300 [4:04:48<10:05:18, 15.85s/it]INFO 04-12 14:51:01 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1010/3300 [4:05:03<9:47:00, 15.38s/it] 2026-04-12 14:51:15,166 | INFO | train_grpo_train | metrics_logged mode=train step=1010

                                                       
{'loss': -0.1126, 'grad_norm': 1.9073106050491333, 'learning_rate': 6.942424242424243e-06, 'num_tokens': 2047701.0, 'completions/mean_length': 167.25, 'completions/min_length': 46.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 52.333335876464844, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.6894205808639526, 'rewards/meter/std': 0.3634391129016876, '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/lexical_plausibility/mean': 0.9311843514442444, 'rewards/lexical_plausibility/std': 0.12742556631565094, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.1505940705537796, 'rewards/repeat_penalty/mean': 0.9380592107772827, 'rewards/repeat_penalty/std': 0.0699658989906311, 'rewards/near_duplicate_penalty/mean': 0.9716432094573975, 'rewards/near_duplicate_penalty/std': 0.01796828769147396, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9655306339263916, 'rewards/distinct_2/std': 0.07098367065191269, 'rewards/total_composite/mean': 0.23814599215984344, 'rewards/total_composite/std': 0.16449278593063354, 'reward': 0.23814599215984344, 'reward_std': 0.16449278593063354, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10613491386175156, 'sampling/sampling_logp_difference/max': 1.6188807487487793, 'sampling/importance_sampling_ratio/min': 0.1981203258037567, 'sampling/importance_sampling_ratio/mean': 1.0171749591827393, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5171218290925026, 'clip_ratio/low_mean': 0.0078125, 'clip_ratio/low_min': 0.0078125, 'clip_ratio/high_mean': 0.05042328080162406, 'clip_ratio/high_max': 0.05042328080162406, 'clip_ratio/region_mean': 0.05823578080162406, 'reward_total_mean': 0.23814599215984344, 'reward_meter_mean': 0.6894205808639526, 'reward_meter_std': 0.3634391129016876, '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_lexical_plausibility_mean': 0.9311843514442444, 'reward_lexical_plausibility_std': 0.12742556631565094, 'reward_repeat_penalty_mean': 0.9380592107772827, 'reward_repeat_penalty_std': 0.0699658989906311, 'reward_near_duplicate_penalty_mean': 0.9716432094573975, 'reward_near_duplicate_penalty_std': 0.01796828769147396, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9655306339263916, 'reward_distinct_2_std': 0.07098367065191269, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.1505940705537796, 'reward_total_composite_mean': 0.23814599215984344, 'reward_total_composite_std': 0.16449278593063354, 'epoch': 0.04}

 31%|███       | 1010/3300 [4:05:03<9:47:00, 15.38s/it]INFO 04-12 14:51:15 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1011/3300 [4:05:16<9:28:59, 14.91s/it]2026-04-12 14:51:28,998 | INFO | train_grpo_train | metrics_logged mode=train step=1011

                                                       
{'loss': -0.0656, 'grad_norm': 3.996352434158325, 'learning_rate': 6.93939393939394e-06, 'num_tokens': 2049358.0, 'completions/mean_length': 109.125, 'completions/min_length': 44.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 51.57143020629883, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.7323673963546753, 'rewards/meter/std': 0.3175550401210785, '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/lexical_plausibility/mean': 0.9644086360931396, 'rewards/lexical_plausibility/std': 0.1006675586104393, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.2133910059928894, 'rewards/repeat_penalty/mean': 0.750518798828125, 'rewards/repeat_penalty/std': 0.3046644628047943, 'rewards/near_duplicate_penalty/mean': 0.9092114567756653, 'rewards/near_duplicate_penalty/std': 0.06783723831176758, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.8898155689239502, 'rewards/distinct_2/std': 0.14818856120109558, 'rewards/total_composite/mean': 0.327912300825119, 'rewards/total_composite/std': 0.21722827851772308, 'reward': 0.327912300825119, 'reward_std': 0.21722827851772308, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1013401448726654, 'sampling/sampling_logp_difference/max': 1.9297844171524048, 'sampling/importance_sampling_ratio/min': 0.14517949521541595, 'sampling/importance_sampling_ratio/mean': 1.0008724927902222, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4322999492287636, 'clip_ratio/low_mean': 0.0066964286379516125, 'clip_ratio/low_min': 0.0066964286379516125, 'clip_ratio/high_mean': 0.07299076020717621, 'clip_ratio/high_max': 0.07299076020717621, 'clip_ratio/region_mean': 0.07968718884512782, 'reward_total_mean': 0.327912300825119, 'reward_meter_mean': 0.7323673963546753, 'reward_meter_std': 0.3175550401210785, '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_lexical_plausibility_mean': 0.9644086360931396, 'reward_lexical_plausibility_std': 0.1006675586104393, 'reward_repeat_penalty_mean': 0.750518798828125, 'reward_repeat_penalty_std': 0.3046644628047943, 'reward_near_duplicate_penalty_mean': 0.9092114567756653, 'reward_near_duplicate_penalty_std': 0.06783723831176758, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.8898155689239502, 'reward_distinct_2_std': 0.14818856120109558, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.2133910059928894, 'reward_total_composite_mean': 0.327912300825119, 'reward_total_composite_std': 0.21722827851772308, 'epoch': 0.04}

 31%|███       | 1011/3300 [4:05:16<9:28:59, 14.91s/it]INFO 04-12 14:51:29 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1012/3300 [4:05:26<8:25:11, 13.25s/it]2026-04-12 14:51:38,348 | INFO | train_grpo_train | metrics_logged mode=train step=1012

                                                       
{'loss': -0.0083, 'grad_norm': 10.422324180603027, 'learning_rate': 6.936363636363636e-06, 'num_tokens': 2051407.0, 'completions/mean_length': 84.125, 'completions/min_length': 59.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 84.125, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.6902055144309998, 'rewards/meter/std': 0.314127653837204, '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/lexical_plausibility/mean': 0.9977678656578064, 'rewards/lexical_plausibility/std': 0.006313450634479523, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.07440237700939178, 'rewards/repeat_penalty/mean': 0.7271653413772583, 'rewards/repeat_penalty/std': 0.2486673891544342, 'rewards/near_duplicate_penalty/mean': 0.9066400527954102, 'rewards/near_duplicate_penalty/std': 0.07484584301710129, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.84401535987854, 'rewards/distinct_2/std': 0.14852824807167053, 'rewards/total_composite/mean': 0.20311975479125977, 'rewards/total_composite/std': 0.11119531095027924, 'reward': 0.20311975479125977, 'reward_std': 0.11119531095027924, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10833962261676788, 'sampling/sampling_logp_difference/max': 1.4471979141235352, 'sampling/importance_sampling_ratio/min': 0.2352285087108612, 'sampling/importance_sampling_ratio/mean': 1.011983871459961, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5167733803391457, 'clip_ratio/low_mean': 0.0486882459372282, 'clip_ratio/low_min': 0.0486882459372282, 'clip_ratio/high_mean': 0.042902862187474966, 'clip_ratio/high_max': 0.042902862187474966, 'clip_ratio/region_mean': 0.09159110812470317, 'reward_total_mean': 0.20311975479125977, 'reward_meter_mean': 0.6902055144309998, 'reward_meter_std': 0.314127653837204, '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_lexical_plausibility_mean': 0.9977678656578064, 'reward_lexical_plausibility_std': 0.006313450634479523, 'reward_repeat_penalty_mean': 0.7271653413772583, 'reward_repeat_penalty_std': 0.2486673891544342, 'reward_near_duplicate_penalty_mean': 0.9066400527954102, 'reward_near_duplicate_penalty_std': 0.07484584301710129, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.84401535987854, 'reward_distinct_2_std': 0.14852824807167053, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.07440237700939178, 'reward_total_composite_mean': 0.20311975479125977, 'reward_total_composite_std': 0.11119531095027924, 'epoch': 0.04}

 31%|███       | 1012/3300 [4:05:26<8:25:11, 13.25s/it]INFO 04-12 14:51:38 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1013/3300 [4:05:35<7:36:51, 11.99s/it]2026-04-12 14:51:47,396 | INFO | train_grpo_train | metrics_logged mode=train step=1013

                                                       
{'loss': 0.0269, 'grad_norm': 9.366815567016602, 'learning_rate': 6.9333333333333344e-06, 'num_tokens': 2053473.0, 'completions/mean_length': 90.25, 'completions/min_length': 83.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 90.25, 'completions/min_terminated_length': 83.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.7935206890106201, 'rewards/meter/std': 0.1681874543428421, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.1927071362733841, 'rewards/repeat_penalty/std': 0.0855286493897438, 'rewards/near_duplicate_penalty/mean': 0.792396605014801, 'rewards/near_duplicate_penalty/std': 0.0862075611948967, 'rewards/opening_diversity/mean': 0.34375, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.6951733231544495, 'rewards/distinct_2/std': 0.0892522931098938, 'rewards/total_composite/mean': 0.277732253074646, 'rewards/total_composite/std': 0.05886560678482056, 'reward': 0.277732253074646, 'reward_std': 0.058865610510110855, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08801070600748062, 'sampling/sampling_logp_difference/max': 1.9264554977416992, 'sampling/importance_sampling_ratio/min': 0.14566358923912048, 'sampling/importance_sampling_ratio/mean': 0.9931285381317139, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4531378522515297, 'clip_ratio/low_mean': 0.046969751827418804, 'clip_ratio/low_min': 0.046969751827418804, 'clip_ratio/high_mean': 0.03944897651672363, 'clip_ratio/high_max': 0.03944897651672363, 'clip_ratio/region_mean': 0.08641872834414244, 'reward_total_mean': 0.277732253074646, 'reward_meter_mean': 0.7935206890106201, 'reward_meter_std': 0.1681874543428421, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.1927071362733841, 'reward_repeat_penalty_std': 0.0855286493897438, 'reward_near_duplicate_penalty_mean': 0.792396605014801, 'reward_near_duplicate_penalty_std': 0.0862075611948967, 'reward_opening_diversity_mean': 0.34375, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.6951733231544495, 'reward_distinct_2_std': 0.0892522931098938, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.277732253074646, 'reward_total_composite_std': 0.05886560678482056, 'epoch': 0.04}

 31%|███       | 1013/3300 [4:05:35<7:36:51, 11.99s/it]INFO 04-12 14:51:47 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1014/3300 [4:05:48<7:54:24, 12.45s/it]2026-04-12 14:52:00,927 | INFO | train_grpo_train | metrics_logged mode=train step=1014

                                                       
{'loss': -0.0594, 'grad_norm': 2.3424248695373535, 'learning_rate': 6.930303030303031e-06, 'num_tokens': 2055155.0, 'completions/mean_length': 235.25, 'completions/min_length': 63.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 69.20000457763672, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.4517163932323456, 'rewards/meter/std': 0.42652854323387146, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.8778268694877625, 'rewards/lexical_plausibility/std': 0.13646115362644196, 'rewards/judge_quality/mean': 0.125, 'rewards/judge_quality/std': 0.0707106813788414, 'rewards/repeat_penalty/mean': 0.7431102991104126, 'rewards/repeat_penalty/std': 0.24545390903949738, 'rewards/near_duplicate_penalty/mean': 0.8971248865127563, 'rewards/near_duplicate_penalty/std': 0.12238641828298569, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.838614821434021, 'rewards/distinct_2/std': 0.21892930567264557, 'rewards/total_composite/mean': 0.035518694669008255, 'rewards/total_composite/std': 0.08521462231874466, 'reward': 0.035518694669008255, 'reward_std': 0.08521462231874466, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10896256566047668, 'sampling/sampling_logp_difference/max': 1.5819063186645508, 'sampling/importance_sampling_ratio/min': 0.20558282732963562, 'sampling/importance_sampling_ratio/mean': 1.0058146715164185, 'sampling/importance_sampling_ratio/max': 1.9902769327163696, 'entropy': 0.3846058249473572, 'clip_ratio/low_mean': 0.062063612043857574, 'clip_ratio/low_min': 0.062063612043857574, 'clip_ratio/high_mean': 0.009493670426309109, 'clip_ratio/high_max': 0.009493670426309109, 'clip_ratio/region_mean': 0.07155728247016668, 'reward_total_mean': 0.035518694669008255, 'reward_meter_mean': 0.4517163932323456, 'reward_meter_std': 0.42652854323387146, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.8778268694877625, 'reward_lexical_plausibility_std': 0.13646115362644196, 'reward_repeat_penalty_mean': 0.7431102991104126, 'reward_repeat_penalty_std': 0.24545390903949738, 'reward_near_duplicate_penalty_mean': 0.8971248865127563, 'reward_near_duplicate_penalty_std': 0.12238641828298569, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.838614821434021, 'reward_distinct_2_std': 0.21892930567264557, 'reward_judge_quality_mean': 0.125, 'reward_judge_quality_std': 0.0707106813788414, 'reward_total_composite_mean': 0.035518694669008255, 'reward_total_composite_std': 0.08521462231874466, 'epoch': 0.04}

 31%|███       | 1014/3300 [4:05:48<7:54:24, 12.45s/it]INFO 04-12 14:52:01 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1015/3300 [4:05:56<6:59:59, 11.03s/it]2026-04-12 14:52:08,638 | INFO | train_grpo_train | metrics_logged mode=train step=1015

                                                       
{'loss': 0.0297, 'grad_norm': 29.638776779174805, 'learning_rate': 6.927272727272728e-06, 'num_tokens': 2056598.0, 'completions/mean_length': 29.375, 'completions/min_length': 26.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.375, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.5122090578079224, 'rewards/meter/std': 0.3878419101238251, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4087499976158142, 'rewards/judge_quality/std': 0.24121345579624176, 'rewards/repeat_penalty/mean': 0.7089154720306396, 'rewards/repeat_penalty/std': 0.06865794956684113, 'rewards/near_duplicate_penalty/mean': 0.9577081203460693, 'rewards/near_duplicate_penalty/std': 0.06238333880901337, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9852070808410645, 'rewards/distinct_2/std': 0.041840631514787674, 'rewards/total_composite/mean': 0.22760291397571564, 'rewards/total_composite/std': 0.2543333172798157, 'reward': 0.22760291397571564, 'reward_std': 0.2543333172798157, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11825455725193024, 'sampling/sampling_logp_difference/max': 1.5048630237579346, 'sampling/importance_sampling_ratio/min': 0.22204771637916565, 'sampling/importance_sampling_ratio/mean': 1.0244112014770508, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.630183719098568, 'clip_ratio/low_mean': 0.08032241277396679, 'clip_ratio/low_min': 0.08032241277396679, 'clip_ratio/high_mean': 0.04841570556163788, 'clip_ratio/high_max': 0.04841570556163788, 'clip_ratio/region_mean': 0.12873811833560467, 'reward_total_mean': 0.22760291397571564, 'reward_meter_mean': 0.5122090578079224, 'reward_meter_std': 0.3878419101238251, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7089154720306396, 'reward_repeat_penalty_std': 0.06865794956684113, 'reward_near_duplicate_penalty_mean': 0.9577081203460693, 'reward_near_duplicate_penalty_std': 0.06238333880901337, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9852070808410645, 'reward_distinct_2_std': 0.041840631514787674, 'reward_judge_quality_mean': 0.4087499976158142, 'reward_judge_quality_std': 0.24121345579624176, 'reward_total_composite_mean': 0.22760291397571564, 'reward_total_composite_std': 0.2543333172798157, 'epoch': 0.04}

 31%|███       | 1015/3300 [4:05:56<6:59:59, 11.03s/it]INFO 04-12 14:52:08 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1016/3300 [4:06:11<7:41:33, 12.12s/it]2026-04-12 14:52:23,327 | INFO | train_grpo_train | metrics_logged mode=train step=1016

                                                       
{'loss': 0.0437, 'grad_norm': 0.2444995641708374, 'learning_rate': 6.9242424242424245e-06, 'num_tokens': 2058807.0, 'completions/mean_length': 495.125, 'completions/min_length': 377.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.875, 'completions/mean_terminated_length': 377.0, 'completions/min_terminated_length': 377.0, 'completions/max_terminated_length': 377.0, 'rewards/meter/mean': 0.5891023874282837, 'rewards/meter/std': 0.4049077332019806, 'rewards/count_adherence/mean': 0.38333332538604736, 'rewards/count_adherence/std': 0.13685818016529083, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.125, 'rewards/judge_quality/std': 0.046291008591651917, 'rewards/repeat_penalty/mean': 0.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 0.37742453813552856, 'rewards/near_duplicate_penalty/std': 0.17442139983177185, 'rewards/opening_diversity/mean': 0.3317132592201233, 'rewards/opening_diversity/std': 0.2342468500137329, 'rewards/distinct_2/mean': 0.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.02793014608323574, 'rewards/total_composite/std': 0.02007874846458435, 'reward': 0.02793014608323574, 'reward_std': 0.0200787466019392, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.029305048286914825, 'sampling/sampling_logp_difference/max': 0.8253111839294434, 'sampling/importance_sampling_ratio/min': 0.5166441202163696, 'sampling/importance_sampling_ratio/mean': 1.0093629360198975, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.01212412491440773, 'clip_ratio/low_mean': 0.0039787800051271915, 'clip_ratio/low_min': 0.0039787800051271915, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0039787800051271915, 'reward_total_mean': 0.02793014608323574, 'reward_meter_mean': 0.5891023874282837, 'reward_meter_std': 0.4049077332019806, 'reward_count_adherence_mean': 0.38333332538604736, 'reward_count_adherence_std': 0.13685818016529083, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.0, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 0.37742453813552856, 'reward_near_duplicate_penalty_std': 0.17442139983177185, 'reward_opening_diversity_mean': 0.3317132592201233, 'reward_opening_diversity_std': 0.2342468500137329, 'reward_distinct_2_mean': 0.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.125, 'reward_judge_quality_std': 0.046291008591651917, 'reward_total_composite_mean': 0.02793014608323574, 'reward_total_composite_std': 0.02007874846458435, 'epoch': 0.04}

 31%|███       | 1016/3300 [4:06:11<7:41:33, 12.12s/it]INFO 04-12 14:52:23 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1017/3300 [4:06:24<7:56:37, 12.53s/it]2026-04-12 14:52:36,787 | INFO | train_grpo_train | metrics_logged mode=train step=1017

                                                       
{'loss': -0.0729, 'grad_norm': 4.517416954040527, 'learning_rate': 6.921212121212122e-06, 'num_tokens': 2060563.0, 'completions/mean_length': 116.5, 'completions/min_length': 52.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 60.000003814697266, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.5273847579956055, 'rewards/meter/std': 0.40735289454460144, '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/lexical_plausibility/mean': 0.962093710899353, 'rewards/lexical_plausibility/std': 0.10721520334482193, 'rewards/judge_quality/mean': 0.2875000238418579, 'rewards/judge_quality/std': 0.1505940705537796, 'rewards/repeat_penalty/mean': 0.6538918614387512, 'rewards/repeat_penalty/std': 0.33201339840888977, 'rewards/near_duplicate_penalty/mean': 0.8698864579200745, 'rewards/near_duplicate_penalty/std': 0.1236475259065628, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.7783924341201782, 'rewards/distinct_2/std': 0.21051128208637238, 'rewards/total_composite/mean': 0.18891584873199463, 'rewards/total_composite/std': 0.16921192407608032, 'reward': 0.18891584873199463, 'reward_std': 0.16921192407608032, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10032311081886292, 'sampling/sampling_logp_difference/max': 2.389025926589966, 'sampling/importance_sampling_ratio/min': 0.09171898663043976, 'sampling/importance_sampling_ratio/mean': 1.007766604423523, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5222587697207928, 'clip_ratio/low_mean': 0.038829339668154716, 'clip_ratio/low_min': 0.038829339668154716, 'clip_ratio/high_mean': 0.04004096332937479, 'clip_ratio/high_max': 0.04004096332937479, 'clip_ratio/region_mean': 0.0788703029975295, 'reward_total_mean': 0.18891584873199463, 'reward_meter_mean': 0.5273847579956055, 'reward_meter_std': 0.40735289454460144, '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_lexical_plausibility_mean': 0.962093710899353, 'reward_lexical_plausibility_std': 0.10721520334482193, 'reward_repeat_penalty_mean': 0.6538918614387512, 'reward_repeat_penalty_std': 0.33201339840888977, 'reward_near_duplicate_penalty_mean': 0.8698864579200745, 'reward_near_duplicate_penalty_std': 0.1236475259065628, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.7783924341201782, 'reward_distinct_2_std': 0.21051128208637238, 'reward_judge_quality_mean': 0.2875000238418579, 'reward_judge_quality_std': 0.1505940705537796, 'reward_total_composite_mean': 0.18891584873199463, 'reward_total_composite_std': 0.16921192407608032, 'epoch': 0.04}

 31%|███       | 1017/3300 [4:06:24<7:56:37, 12.53s/it]INFO 04-12 14:52:37 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1018/3300 [4:06:32<7:03:45, 11.14s/it]2026-04-12 14:52:44,695 | INFO | train_grpo_train | metrics_logged mode=train step=1018

                                                       
{'loss': -0.0295, 'grad_norm': 17.33719825744629, 'learning_rate': 6.918181818181818e-06, 'num_tokens': 2062195.0, 'completions/mean_length': 47.0, 'completions/min_length': 33.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.0, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.7240685224533081, 'rewards/meter/std': 0.33690762519836426, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.3644633889198303, 'rewards/repeat_penalty/std': 0.20901869237422943, 'rewards/near_duplicate_penalty/mean': 0.8008079528808594, 'rewards/near_duplicate_penalty/std': 0.12227211892604828, 'rewards/opening_diversity/mean': 0.671875, 'rewards/opening_diversity/std': 0.14846687018871307, 'rewards/distinct_2/mean': 0.6436425447463989, 'rewards/distinct_2/std': 0.17683729529380798, 'rewards/total_composite/mean': 0.19473877549171448, 'rewards/total_composite/std': 0.11701270192861557, 'reward': 0.19473877549171448, 'reward_std': 0.11701269447803497, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09906559437513351, 'sampling/sampling_logp_difference/max': 1.7095346450805664, 'sampling/importance_sampling_ratio/min': 0.18094998598098755, 'sampling/importance_sampling_ratio/mean': 1.0146312713623047, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6318059228360653, 'clip_ratio/low_mean': 0.04329847078770399, 'clip_ratio/low_min': 0.04329847078770399, 'clip_ratio/high_mean': 0.053069054149091244, 'clip_ratio/high_max': 0.053069054149091244, 'clip_ratio/region_mean': 0.09636752493679523, 'reward_total_mean': 0.19473877549171448, 'reward_meter_mean': 0.7240685224533081, 'reward_meter_std': 0.33690762519836426, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.3644633889198303, 'reward_repeat_penalty_std': 0.20901869237422943, 'reward_near_duplicate_penalty_mean': 0.8008079528808594, 'reward_near_duplicate_penalty_std': 0.12227211892604828, 'reward_opening_diversity_mean': 0.671875, 'reward_opening_diversity_std': 0.14846687018871307, 'reward_distinct_2_mean': 0.6436425447463989, 'reward_distinct_2_std': 0.17683729529380798, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.19473877549171448, 'reward_total_composite_std': 0.11701270192861557, 'epoch': 0.04}

 31%|███       | 1018/3300 [4:06:32<7:03:45, 11.14s/it]INFO 04-12 14:52:44 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1019/3300 [4:06:41<6:34:46, 10.38s/it]2026-04-12 14:52:53,307 | INFO | train_grpo_train | metrics_logged mode=train step=1019

                                                       
{'loss': 0.0571, 'grad_norm': 9.505874633789062, 'learning_rate': 6.915151515151515e-06, 'num_tokens': 2063936.0, 'completions/mean_length': 61.625, 'completions/min_length': 54.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.625, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.7488036155700684, 'rewards/meter/std': 0.2799701690673828, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.2980732023715973, 'rewards/repeat_penalty/std': 0.1365506500005722, 'rewards/near_duplicate_penalty/mean': 0.7738947868347168, 'rewards/near_duplicate_penalty/std': 0.13323625922203064, 'rewards/opening_diversity/mean': 0.515625, 'rewards/opening_diversity/std': 0.19408094882965088, 'rewards/distinct_2/mean': 0.7420055866241455, 'rewards/distinct_2/std': 0.18532700836658478, 'rewards/total_composite/mean': 0.18353229761123657, 'rewards/total_composite/std': 0.1281534880399704, 'reward': 0.18353229761123657, 'reward_std': 0.1281534880399704, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08674348890781403, 'sampling/sampling_logp_difference/max': 1.525856852531433, 'sampling/importance_sampling_ratio/min': 0.21743467450141907, 'sampling/importance_sampling_ratio/mean': 1.0050889253616333, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4259168542921543, 'clip_ratio/low_mean': 0.03343175002373755, 'clip_ratio/low_min': 0.03343175002373755, 'clip_ratio/high_mean': 0.0341534954495728, 'clip_ratio/high_max': 0.0341534954495728, 'clip_ratio/region_mean': 0.06758524547331035, 'reward_total_mean': 0.18353229761123657, 'reward_meter_mean': 0.7488036155700684, 'reward_meter_std': 0.2799701690673828, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.2980732023715973, 'reward_repeat_penalty_std': 0.1365506500005722, 'reward_near_duplicate_penalty_mean': 0.7738947868347168, 'reward_near_duplicate_penalty_std': 0.13323625922203064, 'reward_opening_diversity_mean': 0.515625, 'reward_opening_diversity_std': 0.19408094882965088, 'reward_distinct_2_mean': 0.7420055866241455, 'reward_distinct_2_std': 0.18532700836658478, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.18353229761123657, 'reward_total_composite_std': 0.1281534880399704, 'epoch': 0.04}

 31%|███       | 1019/3300 [4:06:41<6:34:46, 10.38s/it]INFO 04-12 14:52:53 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1020/3300 [4:06:54<7:08:17, 11.27s/it]2026-04-12 14:53:06,652 | INFO | train_grpo_train | metrics_logged mode=train step=1020

                                                       
{'loss': -0.1716, 'grad_norm': 3.393901824951172, 'learning_rate': 6.912121212121212e-06, 'num_tokens': 2066177.0, 'completions/mean_length': 155.125, 'completions/min_length': 99.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 104.14286041259766, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 112.0, 'rewards/meter/mean': 0.8688631057739258, 'rewards/meter/std': 0.2380598783493042, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9745819568634033, 'rewards/lexical_plausibility/std': 0.07189306616783142, 'rewards/judge_quality/mean': 0.13750000298023224, 'rewards/judge_quality/std': 0.0353553406894207, 'rewards/repeat_penalty/mean': 0.2582615315914154, 'rewards/repeat_penalty/std': 0.26931893825531006, 'rewards/near_duplicate_penalty/mean': 0.668350338935852, 'rewards/near_duplicate_penalty/std': 0.1923406720161438, 'rewards/opening_diversity/mean': 0.734375, 'rewards/opening_diversity/std': 0.2868976593017578, 'rewards/distinct_2/mean': 0.47387489676475525, 'rewards/distinct_2/std': 0.3142240047454834, 'rewards/total_composite/mean': 0.11387498676776886, 'rewards/total_composite/std': 0.0582408607006073, 'reward': 0.11387498676776886, 'reward_std': 0.0582408607006073, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07800491899251938, 'sampling/sampling_logp_difference/max': 1.5665688514709473, 'sampling/importance_sampling_ratio/min': 0.20876023173332214, 'sampling/importance_sampling_ratio/mean': 1.01407790184021, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4713069386780262, 'clip_ratio/low_mean': 0.010101010091602802, 'clip_ratio/low_min': 0.010101010091602802, 'clip_ratio/high_mean': 0.06296353740617633, 'clip_ratio/high_max': 0.06296353740617633, 'clip_ratio/region_mean': 0.07306454749777913, 'reward_total_mean': 0.11387498676776886, 'reward_meter_mean': 0.8688631057739258, 'reward_meter_std': 0.2380598783493042, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9745819568634033, 'reward_lexical_plausibility_std': 0.07189306616783142, 'reward_repeat_penalty_mean': 0.2582615315914154, 'reward_repeat_penalty_std': 0.26931893825531006, 'reward_near_duplicate_penalty_mean': 0.668350338935852, 'reward_near_duplicate_penalty_std': 0.1923406720161438, 'reward_opening_diversity_mean': 0.734375, 'reward_opening_diversity_std': 0.2868976593017578, 'reward_distinct_2_mean': 0.47387489676475525, 'reward_distinct_2_std': 0.3142240047454834, 'reward_judge_quality_mean': 0.13750000298023224, 'reward_judge_quality_std': 0.0353553406894207, 'reward_total_composite_mean': 0.11387498676776886, 'reward_total_composite_std': 0.0582408607006073, 'epoch': 0.04}

 31%|███       | 1020/3300 [4:06:54<7:08:17, 11.27s/it]INFO 04-12 14:53:06 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1021/3300 [4:07:08<7:40:21, 12.12s/it]2026-04-12 14:53:20,753 | INFO | train_grpo_train | metrics_logged mode=train step=1021

                                                       
{'loss': -0.0346, 'grad_norm': 4.435267448425293, 'learning_rate': 6.90909090909091e-06, 'num_tokens': 2067755.0, 'completions/mean_length': 112.25, 'completions/min_length': 49.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 55.142860412597656, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.23766043782234192, 'rewards/meter/std': 0.26641353964805603, '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/lexical_plausibility/mean': 0.9752229452133179, 'rewards/lexical_plausibility/std': 0.07008011639118195, 'rewards/judge_quality/mean': 0.21250000596046448, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.6807931661605835, 'rewards/repeat_penalty/std': 0.11142243444919586, 'rewards/near_duplicate_penalty/mean': 0.846206545829773, 'rewards/near_duplicate_penalty/std': 0.0366097055375576, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.8587462902069092, 'rewards/distinct_2/std': 0.07720057666301727, 'rewards/total_composite/mean': 0.052861981093883514, 'rewards/total_composite/std': 0.06511695683002472, 'reward': 0.052861981093883514, 'reward_std': 0.06511695683002472, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0928114801645279, 'sampling/sampling_logp_difference/max': 1.0159813165664673, 'sampling/importance_sampling_ratio/min': 0.3620469868183136, 'sampling/importance_sampling_ratio/mean': 1.0031670331954956, 'sampling/importance_sampling_ratio/max': 1.793241262435913, 'entropy': 0.5795053467154503, 'clip_ratio/low_mean': 0.04669883707538247, 'clip_ratio/low_min': 0.04669883707538247, 'clip_ratio/high_mean': 0.040712181478738785, 'clip_ratio/high_max': 0.040712181478738785, 'clip_ratio/region_mean': 0.08741101855412126, 'reward_total_mean': 0.052861981093883514, 'reward_meter_mean': 0.23766043782234192, 'reward_meter_std': 0.26641353964805603, '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_lexical_plausibility_mean': 0.9752229452133179, 'reward_lexical_plausibility_std': 0.07008011639118195, 'reward_repeat_penalty_mean': 0.6807931661605835, 'reward_repeat_penalty_std': 0.11142243444919586, 'reward_near_duplicate_penalty_mean': 0.846206545829773, 'reward_near_duplicate_penalty_std': 0.0366097055375576, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.8587462902069092, 'reward_distinct_2_std': 0.07720057666301727, 'reward_judge_quality_mean': 0.21250000596046448, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.052861981093883514, 'reward_total_composite_std': 0.06511695683002472, 'epoch': 0.04}

 31%|███       | 1021/3300 [4:07:08<7:40:21, 12.12s/it]INFO 04-12 14:53:20 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1022/3300 [4:07:22<8:01:08, 12.67s/it]2026-04-12 14:53:34,715 | INFO | train_grpo_train | metrics_logged mode=train step=1022

                                                       
{'loss': -0.1106, 'grad_norm': 2.7205684185028076, 'learning_rate': 6.906060606060606e-06, 'num_tokens': 2069446.0, 'completions/mean_length': 176.375, 'completions/min_length': 58.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 64.5, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.6154673099517822, 'rewards/meter/std': 0.29473501443862915, '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/lexical_plausibility/mean': 0.9299918413162231, 'rewards/lexical_plausibility/std': 0.12978941202163696, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.7901404500007629, 'rewards/repeat_penalty/std': 0.2279074341058731, 'rewards/near_duplicate_penalty/mean': 0.9328340291976929, 'rewards/near_duplicate_penalty/std': 0.06356509774923325, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.8766018152236938, 'rewards/distinct_2/std': 0.11375460028648376, 'rewards/total_composite/mean': 0.14203496277332306, 'rewards/total_composite/std': 0.1235874742269516, 'reward': 0.14203496277332306, 'reward_std': 0.1235874742269516, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09047681838274002, 'sampling/sampling_logp_difference/max': 1.426063060760498, 'sampling/importance_sampling_ratio/min': 0.24025292694568634, 'sampling/importance_sampling_ratio/mean': 1.0126465559005737, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.42599399015307426, 'clip_ratio/low_mean': 0.01293103490024805, 'clip_ratio/low_min': 0.01293103490024805, 'clip_ratio/high_mean': 0.049724617041647434, 'clip_ratio/high_max': 0.049724617041647434, 'clip_ratio/region_mean': 0.06265565194189548, 'reward_total_mean': 0.14203496277332306, 'reward_meter_mean': 0.6154673099517822, 'reward_meter_std': 0.29473501443862915, '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_lexical_plausibility_mean': 0.9299918413162231, 'reward_lexical_plausibility_std': 0.12978941202163696, 'reward_repeat_penalty_mean': 0.7901404500007629, 'reward_repeat_penalty_std': 0.2279074341058731, 'reward_near_duplicate_penalty_mean': 0.9328340291976929, 'reward_near_duplicate_penalty_std': 0.06356509774923325, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.8766018152236938, 'reward_distinct_2_std': 0.11375460028648376, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.14203496277332306, 'reward_total_composite_std': 0.1235874742269516, 'epoch': 0.04}

 31%|███       | 1022/3300 [4:07:22<8:01:08, 12.67s/it]INFO 04-12 14:53:34 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1023/3300 [4:07:36<8:13:11, 13.00s/it]2026-04-12 14:53:48,470 | INFO | train_grpo_train | metrics_logged mode=train step=1023

                                                       
{'loss': -0.076, 'grad_norm': 1.9131355285644531, 'learning_rate': 6.903030303030304e-06, 'num_tokens': 2071076.0, 'completions/mean_length': 225.75, 'completions/min_length': 51.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.17156369984149933, 'rewards/meter/std': 0.1722494512796402, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.9244678020477295, 'rewards/lexical_plausibility/std': 0.12275119870901108, 'rewards/judge_quality/mean': 0.13124999403953552, 'rewards/judge_quality/std': 0.1066954955458641, 'rewards/repeat_penalty/mean': 0.6404745578765869, 'rewards/repeat_penalty/std': 0.22385944426059723, 'rewards/near_duplicate_penalty/mean': 0.7799478769302368, 'rewards/near_duplicate_penalty/std': 0.11074870824813843, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.8301138877868652, 'rewards/distinct_2/std': 0.1930258572101593, 'rewards/total_composite/mean': 0.031589772552251816, 'rewards/total_composite/std': 0.03598390519618988, 'reward': 0.031589772552251816, 'reward_std': 0.03598390147089958, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1007244810461998, 'sampling/sampling_logp_difference/max': 1.2299976348876953, 'sampling/importance_sampling_ratio/min': 0.2922932505607605, 'sampling/importance_sampling_ratio/mean': 1.0003584623336792, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3576190359890461, 'clip_ratio/low_mean': 0.013636363670229912, 'clip_ratio/low_min': 0.013636363670229912, 'clip_ratio/high_mean': 0.04884940106421709, 'clip_ratio/high_max': 0.04884940106421709, 'clip_ratio/region_mean': 0.062485764734447, 'reward_total_mean': 0.031589772552251816, 'reward_meter_mean': 0.17156369984149933, 'reward_meter_std': 0.1722494512796402, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.9244678020477295, 'reward_lexical_plausibility_std': 0.12275119870901108, 'reward_repeat_penalty_mean': 0.6404745578765869, 'reward_repeat_penalty_std': 0.22385944426059723, 'reward_near_duplicate_penalty_mean': 0.7799478769302368, 'reward_near_duplicate_penalty_std': 0.11074870824813843, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.8301138877868652, 'reward_distinct_2_std': 0.1930258572101593, 'reward_judge_quality_mean': 0.13124999403953552, 'reward_judge_quality_std': 0.1066954955458641, 'reward_total_composite_mean': 0.031589772552251816, 'reward_total_composite_std': 0.03598390519618988, 'epoch': 0.04}

 31%|███       | 1023/3300 [4:07:36<8:13:11, 13.00s/it]INFO 04-12 14:53:48 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1024/3300 [4:07:50<8:22:59, 13.26s/it]2026-04-12 14:54:02,345 | INFO | train_grpo_train | metrics_logged mode=train step=1024

                                                       
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 6.9e-06, 'num_tokens': 2072628.0, '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': 0.7472419142723083, 'rewards/meter/std': 0.23165880143642426, 'rewards/count_adherence/mean': 0.8571428656578064, 'rewards/count_adherence/std': 0.2645200192928314, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9739459753036499, 'rewards/lexical_plausibility/std': 0.07369180768728256, 'rewards/judge_quality/mean': 0.07500000298023224, 'rewards/judge_quality/std': 0.046291008591651917, 'rewards/repeat_penalty/mean': 0.037902988493442535, 'rewards/repeat_penalty/std': 0.1072058379650116, 'rewards/near_duplicate_penalty/mean': 0.256704717874527, 'rewards/near_duplicate_penalty/std': 0.2252495139837265, 'rewards/opening_diversity/mean': 0.2873000204563141, 'rewards/opening_diversity/std': 0.28143349289894104, 'rewards/distinct_2/mean': 0.11625874042510986, 'rewards/distinct_2/std': 0.3288293778896332, 'rewards/total_composite/mean': 0.04951168969273567, 'rewards/total_composite/std': 0.04630482941865921, 'reward': 0.04951168969273567, 'reward_std': 0.04630482941865921, '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': 0.04951168969273567, 'reward_meter_mean': 0.7472419142723083, 'reward_meter_std': 0.23165880143642426, 'reward_count_adherence_mean': 0.8571428656578064, 'reward_count_adherence_std': 0.2645200192928314, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9739459753036499, 'reward_lexical_plausibility_std': 0.07369180768728256, 'reward_repeat_penalty_mean': 0.037902988493442535, 'reward_repeat_penalty_std': 0.1072058379650116, 'reward_near_duplicate_penalty_mean': 0.256704717874527, 'reward_near_duplicate_penalty_std': 0.2252495139837265, 'reward_opening_diversity_mean': 0.2873000204563141, 'reward_opening_diversity_std': 0.28143349289894104, 'reward_distinct_2_mean': 0.11625874042510986, 'reward_distinct_2_std': 0.3288293778896332, 'reward_judge_quality_mean': 0.07500000298023224, 'reward_judge_quality_std': 0.046291008591651917, 'reward_total_composite_mean': 0.04951168969273567, 'reward_total_composite_std': 0.04630482941865921, 'epoch': 0.04}

 31%|███       | 1024/3300 [4:07:50<8:22:59, 13.26s/it]INFO 04-12 14:54:02 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1025/3300 [4:08:00<7:44:39, 12.25s/it]2026-04-12 14:54:12,252 | INFO | train_grpo_train | metrics_logged mode=train step=1025

                                                       
{'loss': 0.0041, 'grad_norm': 7.720107078552246, 'learning_rate': 6.896969696969697e-06, 'num_tokens': 2074609.0, 'completions/mean_length': 87.625, 'completions/min_length': 74.0, 'completions/max_length': 118.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 87.625, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 118.0, 'rewards/meter/mean': 0.5218161940574646, 'rewards/meter/std': 0.3966940641403198, '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/lexical_plausibility/mean': 0.9539638757705688, 'rewards/lexical_plausibility/std': 0.06820890307426453, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.11649646610021591, 'rewards/repeat_penalty/mean': 0.7015840411186218, 'rewards/repeat_penalty/std': 0.2907060384750366, 'rewards/near_duplicate_penalty/mean': 0.8775174617767334, 'rewards/near_duplicate_penalty/std': 0.10983667522668839, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.7779926657676697, 'rewards/distinct_2/std': 0.2638917565345764, 'rewards/total_composite/mean': 0.13111929595470428, 'rewards/total_composite/std': 0.09954511374235153, 'reward': 0.13111929595470428, 'reward_std': 0.09954510629177094, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10262478142976761, 'sampling/sampling_logp_difference/max': 1.7458510398864746, 'sampling/importance_sampling_ratio/min': 0.17449642717838287, 'sampling/importance_sampling_ratio/mean': 0.9998663663864136, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5056793540716171, 'clip_ratio/low_mean': 0.05708865029737353, 'clip_ratio/low_min': 0.05708865029737353, 'clip_ratio/high_mean': 0.04222492501139641, 'clip_ratio/high_max': 0.04222492501139641, 'clip_ratio/region_mean': 0.09931357530876994, 'reward_total_mean': 0.13111929595470428, 'reward_meter_mean': 0.5218161940574646, 'reward_meter_std': 0.3966940641403198, '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_lexical_plausibility_mean': 0.9539638757705688, 'reward_lexical_plausibility_std': 0.06820890307426453, 'reward_repeat_penalty_mean': 0.7015840411186218, 'reward_repeat_penalty_std': 0.2907060384750366, 'reward_near_duplicate_penalty_mean': 0.8775174617767334, 'reward_near_duplicate_penalty_std': 0.10983667522668839, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.7779926657676697, 'reward_distinct_2_std': 0.2638917565345764, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.11649646610021591, 'reward_total_composite_mean': 0.13111929595470428, 'reward_total_composite_std': 0.09954511374235153, 'epoch': 0.04}

 31%|███       | 1025/3300 [4:08:00<7:44:39, 12.25s/it]INFO 04-12 14:54:12 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1026/3300 [4:08:08<7:00:02, 11.08s/it]2026-04-12 14:54:20,604 | INFO | train_grpo_train | metrics_logged mode=train step=1026

                                                       
{'loss': 0.0397, 'grad_norm': 11.36959171295166, 'learning_rate': 6.893939393939395e-06, 'num_tokens': 2076528.0, 'completions/mean_length': 61.875, 'completions/min_length': 54.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.875, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.5908125042915344, 'rewards/meter/std': 0.463847815990448, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.17500001192092896, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/repeat_penalty/mean': 0.6685135960578918, 'rewards/repeat_penalty/std': 0.23664137721061707, 'rewards/near_duplicate_penalty/mean': 0.8620654344558716, 'rewards/near_duplicate_penalty/std': 0.06888401508331299, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.829962432384491, 'rewards/distinct_2/std': 0.1638273000717163, 'rewards/total_composite/mean': 0.08959561586380005, 'rewards/total_composite/std': 0.06829588115215302, 'reward': 0.08959561586380005, 'reward_std': 0.06829588115215302, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08918594568967819, 'sampling/sampling_logp_difference/max': 2.104063034057617, 'sampling/importance_sampling_ratio/min': 0.12195988744497299, 'sampling/importance_sampling_ratio/mean': 1.0066965818405151, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.49786342680454254, 'clip_ratio/low_mean': 0.02265835809521377, 'clip_ratio/low_min': 0.02265835809521377, 'clip_ratio/high_mean': 0.07138597592711449, 'clip_ratio/high_max': 0.07138597592711449, 'clip_ratio/region_mean': 0.09404433402232826, 'reward_total_mean': 0.08959561586380005, 'reward_meter_mean': 0.5908125042915344, 'reward_meter_std': 0.463847815990448, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6685135960578918, 'reward_repeat_penalty_std': 0.23664137721061707, 'reward_near_duplicate_penalty_mean': 0.8620654344558716, 'reward_near_duplicate_penalty_std': 0.06888401508331299, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.829962432384491, 'reward_distinct_2_std': 0.1638273000717163, 'reward_judge_quality_mean': 0.17500001192092896, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.08959561586380005, 'reward_total_composite_std': 0.06829588115215302, 'epoch': 0.04}

 31%|███       | 1026/3300 [4:08:08<7:00:02, 11.08s/it]INFO 04-12 14:54:20 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1027/3300 [4:08:22<7:28:02, 11.83s/it]2026-04-12 14:54:34,167 | INFO | train_grpo_train | metrics_logged mode=train step=1027

                                                       
{'loss': -0.0743, 'grad_norm': 3.760767936706543, 'learning_rate': 6.890909090909092e-06, 'num_tokens': 2078571.0, 'completions/mean_length': 144.375, 'completions/min_length': 81.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 91.85714721679688, 'completions/min_terminated_length': 81.0, 'completions/max_terminated_length': 124.0, 'rewards/meter/mean': 0.8890489339828491, 'rewards/meter/std': 0.13380175828933716, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9246917366981506, 'rewards/lexical_plausibility/std': 0.11682187020778656, 'rewards/judge_quality/mean': 0.17500001192092896, 'rewards/judge_quality/std': 0.11649647355079651, 'rewards/repeat_penalty/mean': 0.20516061782836914, 'rewards/repeat_penalty/std': 0.2608588635921478, 'rewards/near_duplicate_penalty/mean': 0.6434073448181152, 'rewards/near_duplicate_penalty/std': 0.1436055600643158, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.18898223340511322, 'rewards/distinct_2/mean': 0.38242679834365845, 'rewards/distinct_2/std': 0.36785444617271423, 'rewards/total_composite/mean': 0.13922297954559326, 'rewards/total_composite/std': 0.10259076952934265, 'reward': 0.13922297954559326, 'reward_std': 0.10259077697992325, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0748460665345192, 'sampling/sampling_logp_difference/max': 1.680635929107666, 'sampling/importance_sampling_ratio/min': 0.18625548481941223, 'sampling/importance_sampling_ratio/mean': 1.0046155452728271, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.33101135678589344, 'clip_ratio/low_mean': 0.009350667241960764, 'clip_ratio/low_min': 0.009350667241960764, 'clip_ratio/high_mean': 0.04860563948750496, 'clip_ratio/high_max': 0.04860563948750496, 'clip_ratio/region_mean': 0.05795630672946572, 'reward_total_mean': 0.13922297954559326, 'reward_meter_mean': 0.8890489339828491, 'reward_meter_std': 0.13380175828933716, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9246917366981506, 'reward_lexical_plausibility_std': 0.11682187020778656, 'reward_repeat_penalty_mean': 0.20516061782836914, 'reward_repeat_penalty_std': 0.2608588635921478, 'reward_near_duplicate_penalty_mean': 0.6434073448181152, 'reward_near_duplicate_penalty_std': 0.1436055600643158, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.18898223340511322, 'reward_distinct_2_mean': 0.38242679834365845, 'reward_distinct_2_std': 0.36785444617271423, 'reward_judge_quality_mean': 0.17500001192092896, 'reward_judge_quality_std': 0.11649647355079651, 'reward_total_composite_mean': 0.13922297954559326, 'reward_total_composite_std': 0.10259076952934265, 'epoch': 0.04}

 31%|███       | 1027/3300 [4:08:22<7:28:02, 11.83s/it]INFO 04-12 14:54:34 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1028/3300 [4:08:30<6:50:10, 10.83s/it]2026-04-12 14:54:42,669 | INFO | train_grpo_train | metrics_logged mode=train step=1028

                                                       
{'loss': 0.0749, 'grad_norm': 17.88671112060547, 'learning_rate': 6.887878787878789e-06, 'num_tokens': 2080150.0, 'completions/mean_length': 42.375, 'completions/min_length': 36.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.375, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.4214051365852356, 'rewards/meter/std': 0.394690603017807, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9633767008781433, 'rewards/lexical_plausibility/std': 0.1035863608121872, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.15059404075145721, 'rewards/repeat_penalty/mean': 0.8653683662414551, 'rewards/repeat_penalty/std': 0.15469643473625183, 'rewards/near_duplicate_penalty/mean': 0.9038918614387512, 'rewards/near_duplicate_penalty/std': 0.12023397535085678, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.953242838382721, 'rewards/distinct_2/std': 0.06967140734195709, 'rewards/total_composite/mean': 0.15121299028396606, 'rewards/total_composite/std': 0.18406309187412262, 'reward': 0.15121299028396606, 'reward_std': 0.18406307697296143, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16957247257232666, 'sampling/sampling_logp_difference/max': 1.6085577011108398, 'sampling/importance_sampling_ratio/min': 0.20017611980438232, 'sampling/importance_sampling_ratio/mean': 1.0263725519180298, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7976130917668343, 'clip_ratio/low_mean': 0.09178000781685114, 'clip_ratio/low_min': 0.09178000781685114, 'clip_ratio/high_mean': 0.04741337709128857, 'clip_ratio/high_max': 0.04741337709128857, 'clip_ratio/region_mean': 0.1391933849081397, 'reward_total_mean': 0.15121299028396606, 'reward_meter_mean': 0.4214051365852356, 'reward_meter_std': 0.394690603017807, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9633767008781433, 'reward_lexical_plausibility_std': 0.1035863608121872, 'reward_repeat_penalty_mean': 0.8653683662414551, 'reward_repeat_penalty_std': 0.15469643473625183, 'reward_near_duplicate_penalty_mean': 0.9038918614387512, 'reward_near_duplicate_penalty_std': 0.12023397535085678, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.953242838382721, 'reward_distinct_2_std': 0.06967140734195709, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.15059404075145721, 'reward_total_composite_mean': 0.15121299028396606, 'reward_total_composite_std': 0.18406309187412262, 'epoch': 0.04}

 31%|███       | 1028/3300 [4:08:30<6:50:10, 10.83s/it]INFO 04-12 14:54:42 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1029/3300 [4:08:38<6:19:50, 10.04s/it]2026-04-12 14:54:50,850 | INFO | train_grpo_train | metrics_logged mode=train step=1029

                                                       
{'loss': 0.0212, 'grad_norm': 12.958258628845215, 'learning_rate': 6.8848484848484854e-06, 'num_tokens': 2081618.0, 'completions/mean_length': 27.5, 'completions/min_length': 24.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.5, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.7768885493278503, 'rewards/meter/std': 0.30060726404190063, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4675000309944153, 'rewards/judge_quality/std': 0.32056868076324463, 'rewards/repeat_penalty/mean': 0.7310692071914673, 'rewards/repeat_penalty/std': 0.03878411650657654, 'rewards/near_duplicate_penalty/mean': 0.9747589826583862, 'rewards/near_duplicate_penalty/std': 0.051712166517972946, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.38241299986839294, 'rewards/total_composite/std': 0.2848397195339203, 'reward': 0.38241299986839294, 'reward_std': 0.2848396897315979, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11901742219924927, 'sampling/sampling_logp_difference/max': 1.1289299726486206, 'sampling/importance_sampling_ratio/min': 0.3233790993690491, 'sampling/importance_sampling_ratio/mean': 1.0196528434753418, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7172144502401352, 'clip_ratio/low_mean': 0.07219195133075118, 'clip_ratio/low_min': 0.07219195133075118, 'clip_ratio/high_mean': 0.07204362191259861, 'clip_ratio/high_max': 0.07204362191259861, 'clip_ratio/region_mean': 0.1442355732433498, 'reward_total_mean': 0.38241299986839294, 'reward_meter_mean': 0.7768885493278503, 'reward_meter_std': 0.30060726404190063, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7310692071914673, 'reward_repeat_penalty_std': 0.03878411650657654, 'reward_near_duplicate_penalty_mean': 0.9747589826583862, 'reward_near_duplicate_penalty_std': 0.051712166517972946, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4675000309944153, 'reward_judge_quality_std': 0.32056868076324463, 'reward_total_composite_mean': 0.38241299986839294, 'reward_total_composite_std': 0.2848397195339203, 'epoch': 0.04}

 31%|███       | 1029/3300 [4:08:38<6:19:50, 10.04s/it]INFO 04-12 14:54:51 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1030/3300 [4:08:52<6:58:10, 11.05s/it]2026-04-12 14:55:04,276 | INFO | train_grpo_train | metrics_logged mode=train step=1030

                                                       
{'loss': -0.0034, 'grad_norm': 3.967888832092285, 'learning_rate': 6.881818181818183e-06, 'num_tokens': 2083153.0, 'completions/mean_length': 107.875, 'completions/min_length': 43.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 50.142860412597656, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.13707971572875977, 'rewards/meter/std': 0.2192526012659073, '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/lexical_plausibility/mean': 0.968704104423523, 'rewards/lexical_plausibility/std': 0.08851820975542068, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.11649647355079651, 'rewards/repeat_penalty/mean': 0.7831759452819824, 'rewards/repeat_penalty/std': 0.1511973887681961, 'rewards/near_duplicate_penalty/mean': 0.8563976287841797, 'rewards/near_duplicate_penalty/std': 0.07717317342758179, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9078867435455322, 'rewards/distinct_2/std': 0.11104637384414673, 'rewards/total_composite/mean': 0.042914994060993195, 'rewards/total_composite/std': 0.07613904029130936, 'reward': 0.042914994060993195, 'reward_std': 0.07613903284072876, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11119411885738373, 'sampling/sampling_logp_difference/max': 1.961724042892456, 'sampling/importance_sampling_ratio/min': 0.1406157910823822, 'sampling/importance_sampling_ratio/mean': 0.9975777268409729, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5121523290872574, 'clip_ratio/low_mean': 0.06146345101296902, 'clip_ratio/low_min': 0.06146345101296902, 'clip_ratio/high_mean': 0.02526205498725176, 'clip_ratio/high_max': 0.02526205498725176, 'clip_ratio/region_mean': 0.08672550600022078, 'reward_total_mean': 0.042914994060993195, 'reward_meter_mean': 0.13707971572875977, 'reward_meter_std': 0.2192526012659073, '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_lexical_plausibility_mean': 0.968704104423523, 'reward_lexical_plausibility_std': 0.08851820975542068, 'reward_repeat_penalty_mean': 0.7831759452819824, 'reward_repeat_penalty_std': 0.1511973887681961, 'reward_near_duplicate_penalty_mean': 0.8563976287841797, 'reward_near_duplicate_penalty_std': 0.07717317342758179, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9078867435455322, 'reward_distinct_2_std': 0.11104637384414673, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.11649647355079651, 'reward_total_composite_mean': 0.042914994060993195, 'reward_total_composite_std': 0.07613904029130936, 'epoch': 0.04}

 31%|███       | 1030/3300 [4:08:52<6:58:10, 11.05s/it]INFO 04-12 14:55:04 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1031/3300 [4:09:02<6:50:03, 10.84s/it]2026-04-12 14:55:14,678 | INFO | train_grpo_train | metrics_logged mode=train step=1031

                                                       
{'loss': 0.0984, 'grad_norm': 9.080733299255371, 'learning_rate': 6.878787878787879e-06, 'num_tokens': 2085561.0, 'completions/mean_length': 119.0, 'completions/min_length': 95.0, 'completions/max_length': 132.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 119.0, 'completions/min_terminated_length': 95.0, 'completions/max_terminated_length': 132.0, 'rewards/meter/mean': 0.19246211647987366, 'rewards/meter/std': 0.2490117847919464, '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/lexical_plausibility/mean': 0.9461076855659485, 'rewards/lexical_plausibility/std': 0.09910210967063904, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.5193793177604675, 'rewards/repeat_penalty/std': 0.32295694947242737, 'rewards/near_duplicate_penalty/mean': 0.8403446674346924, 'rewards/near_duplicate_penalty/std': 0.14314475655555725, 'rewards/opening_diversity/mean': 0.921875, 'rewards/opening_diversity/std': 0.17598575353622437, 'rewards/distinct_2/mean': 0.6442869901657104, 'rewards/distinct_2/std': 0.27760687470436096, 'rewards/total_composite/mean': 0.039708416908979416, 'rewards/total_composite/std': 0.039796024560928345, 'reward': 0.039708416908979416, 'reward_std': 0.039796024560928345, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.107515349984169, 'sampling/sampling_logp_difference/max': 1.7417142391204834, 'sampling/importance_sampling_ratio/min': 0.17521977424621582, 'sampling/importance_sampling_ratio/mean': 0.9976531267166138, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4511689208447933, 'clip_ratio/low_mean': 0.052274144254624844, 'clip_ratio/low_min': 0.052274144254624844, 'clip_ratio/high_mean': 0.041469668969511986, 'clip_ratio/high_max': 0.041469668969511986, 'clip_ratio/region_mean': 0.09374381322413683, 'reward_total_mean': 0.039708416908979416, 'reward_meter_mean': 0.19246211647987366, 'reward_meter_std': 0.2490117847919464, '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_lexical_plausibility_mean': 0.9461076855659485, 'reward_lexical_plausibility_std': 0.09910210967063904, 'reward_repeat_penalty_mean': 0.5193793177604675, 'reward_repeat_penalty_std': 0.32295694947242737, 'reward_near_duplicate_penalty_mean': 0.8403446674346924, 'reward_near_duplicate_penalty_std': 0.14314475655555725, 'reward_opening_diversity_mean': 0.921875, 'reward_opening_diversity_std': 0.17598575353622437, 'reward_distinct_2_mean': 0.6442869901657104, 'reward_distinct_2_std': 0.27760687470436096, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.039708416908979416, 'reward_total_composite_std': 0.039796024560928345, 'epoch': 0.04}

 31%|███       | 1031/3300 [4:09:02<6:50:03, 10.84s/it]INFO 04-12 14:55:14 [block_pool.py:378] Successfully reset prefix cache

 31%|███▏      | 1032/3300 [4:09:17<7:37:44, 12.11s/it]2026-04-12 14:55:29,694 | INFO | train_grpo_train | metrics_logged mode=train step=1032

                                                       
{'loss': 0.0006, 'grad_norm': 1.7917559146881104, 'learning_rate': 6.875757575757576e-06, 'num_tokens': 2090237.0, 'completions/mean_length': 377.5, 'completions/min_length': 356.0, 'completions/max_length': 400.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 377.5, 'completions/min_terminated_length': 356.0, 'completions/max_terminated_length': 400.0, 'rewards/meter/mean': 0.4655001163482666, 'rewards/meter/std': 0.3577215373516083, 'rewards/count_adherence/mean': 0.4625000059604645, 'rewards/count_adherence/std': 0.1597989946603775, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 0.28539353609085083, 'rewards/near_duplicate_penalty/std': 0.19244235754013062, 'rewards/opening_diversity/mean': 0.2334536910057068, 'rewards/opening_diversity/std': 0.34575384855270386, 'rewards/distinct_2/mean': 0.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.03236306086182594, 'rewards/total_composite/std': 0.029856164008378983, 'reward': 0.03236306086182594, 'reward_std': 0.029856164008378983, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.016814591363072395, 'sampling/sampling_logp_difference/max': 2.3987481594085693, 'sampling/importance_sampling_ratio/min': 0.09083158522844315, 'sampling/importance_sampling_ratio/mean': 1.0003407001495361, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.08910981006920338, 'clip_ratio/low_mean': 0.0055707081919535995, 'clip_ratio/low_min': 0.0055707081919535995, 'clip_ratio/high_mean': 0.009129347803536803, 'clip_ratio/high_max': 0.009129347803536803, 'clip_ratio/region_mean': 0.014700055995490402, 'reward_total_mean': 0.03236306086182594, 'reward_meter_mean': 0.4655001163482666, 'reward_meter_std': 0.3577215373516083, 'reward_count_adherence_mean': 0.4625000059604645, 'reward_count_adherence_std': 0.1597989946603775, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.0, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 0.28539353609085083, 'reward_near_duplicate_penalty_std': 0.19244235754013062, 'reward_opening_diversity_mean': 0.2334536910057068, 'reward_opening_diversity_std': 0.34575384855270386, 'reward_distinct_2_mean': 0.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.03236306086182594, 'reward_total_composite_std': 0.029856164008378983, 'epoch': 0.04}

 31%|███▏      | 1032/3300 [4:09:17<7:37:44, 12.11s/it]INFO 04-12 14:55:29 [block_pool.py:378] Successfully reset prefix cache

 31%|███▏      | 1033/3300 [4:09:27<7:13:02, 11.46s/it]2026-04-12 14:55:39,644 | INFO | train_grpo_train | metrics_logged mode=train step=1033

                                                       
{'loss': 0.0266, 'grad_norm': 15.333620071411133, 'learning_rate': 6.872727272727273e-06, 'num_tokens': 2091828.0, 'completions/mean_length': 38.875, 'completions/min_length': 33.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.875, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.5671983957290649, 'rewards/meter/std': 0.3471923768520355, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4300000071525574, 'rewards/judge_quality/std': 0.32702118158340454, 'rewards/repeat_penalty/mean': 0.954290509223938, 'rewards/repeat_penalty/std': 0.054342806339263916, 'rewards/near_duplicate_penalty/mean': 0.9662845134735107, 'rewards/near_duplicate_penalty/std': 0.03097359836101532, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9871795177459717, 'rewards/distinct_2/std': 0.03626188635826111, 'rewards/total_composite/mean': 0.21451368927955627, 'rewards/total_composite/std': 0.21267272531986237, 'reward': 0.21451368927955627, 'reward_std': 0.21267271041870117, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1342097967863083, 'sampling/sampling_logp_difference/max': 2.7786865234375, 'sampling/importance_sampling_ratio/min': 0.062120046466588974, 'sampling/importance_sampling_ratio/mean': 0.9889594912528992, 'sampling/importance_sampling_ratio/max': 1.9090549945831299, 'entropy': 0.6256684362888336, 'clip_ratio/low_mean': 0.04960537515580654, 'clip_ratio/low_min': 0.04960537515580654, 'clip_ratio/high_mean': 0.08651018049567938, 'clip_ratio/high_max': 0.08651018049567938, 'clip_ratio/region_mean': 0.13611555565148592, 'reward_total_mean': 0.21451368927955627, 'reward_meter_mean': 0.5671983957290649, 'reward_meter_std': 0.3471923768520355, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.954290509223938, 'reward_repeat_penalty_std': 0.054342806339263916, 'reward_near_duplicate_penalty_mean': 0.9662845134735107, 'reward_near_duplicate_penalty_std': 0.03097359836101532, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9871795177459717, 'reward_distinct_2_std': 0.03626188635826111, 'reward_judge_quality_mean': 0.4300000071525574, 'reward_judge_quality_std': 0.32702118158340454, 'reward_total_composite_mean': 0.21451368927955627, 'reward_total_composite_std': 0.21267272531986237, 'epoch': 0.04}

 31%|███▏      | 1033/3300 [4:09:27<7:13:02, 11.46s/it]INFO 04-12 14:55:39 [block_pool.py:378] Successfully reset prefix cache

 31%|███▏      | 1034/3300 [4:09:41<7:38:23, 12.14s/it]2026-04-12 14:55:53,355 | INFO | train_grpo_train | metrics_logged mode=train step=1034

                                                       
{'loss': -0.0362, 'grad_norm': 1.901582956314087, 'learning_rate': 6.869696969696971e-06, 'num_tokens': 2093361.0, 'completions/mean_length': 223.625, 'completions/min_length': 44.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 50.60000228881836, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.1044534295797348, 'rewards/meter/std': 0.13273285329341888, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/lexical_plausibility/mean': 0.8923224210739136, 'rewards/lexical_plausibility/std': 0.13520511984825134, 'rewards/judge_quality/mean': 0.11249999701976776, 'rewards/judge_quality/std': 0.051754921674728394, 'rewards/repeat_penalty/mean': 0.46667566895484924, 'rewards/repeat_penalty/std': 0.38008829951286316, 'rewards/near_duplicate_penalty/mean': 0.7181737422943115, 'rewards/near_duplicate_penalty/std': 0.22814202308654785, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.25877460837364197, 'rewards/distinct_2/mean': 0.7104001045227051, 'rewards/distinct_2/std': 0.2734965682029724, 'rewards/total_composite/mean': 0.009760160930454731, 'rewards/total_composite/std': 0.020280148833990097, 'reward': 0.009760160930454731, 'reward_std': 0.020280148833990097, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08022608608007431, 'sampling/sampling_logp_difference/max': 1.5839738845825195, 'sampling/importance_sampling_ratio/min': 0.20515820384025574, 'sampling/importance_sampling_ratio/mean': 1.0058679580688477, 'sampling/importance_sampling_ratio/max': 1.6019577980041504, 'entropy': 0.31564923003315926, 'clip_ratio/low_mean': 0.04416412580758333, 'clip_ratio/low_min': 0.04416412580758333, 'clip_ratio/high_mean': 0.01785714365541935, 'clip_ratio/high_max': 0.01785714365541935, 'clip_ratio/region_mean': 0.06202126946300268, 'reward_total_mean': 0.009760160930454731, 'reward_meter_mean': 0.1044534295797348, 'reward_meter_std': 0.13273285329341888, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_lexical_plausibility_mean': 0.8923224210739136, 'reward_lexical_plausibility_std': 0.13520511984825134, 'reward_repeat_penalty_mean': 0.46667566895484924, 'reward_repeat_penalty_std': 0.38008829951286316, 'reward_near_duplicate_penalty_mean': 0.7181737422943115, 'reward_near_duplicate_penalty_std': 0.22814202308654785, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.25877460837364197, 'reward_distinct_2_mean': 0.7104001045227051, 'reward_distinct_2_std': 0.2734965682029724, 'reward_judge_quality_mean': 0.11249999701976776, 'reward_judge_quality_std': 0.051754921674728394, 'reward_total_composite_mean': 0.009760160930454731, 'reward_total_composite_std': 0.020280148833990097, 'epoch': 0.04}

 31%|███▏      | 1034/3300 [4:09:41<7:38:23, 12.14s/it]INFO 04-12 14:55:53 [block_pool.py:378] Successfully reset prefix cache

 31%|███▏      | 1035/3300 [4:09:54<7:47:51, 12.39s/it]2026-04-12 14:56:06,354 | INFO | train_grpo_train | metrics_logged mode=train step=1035

                                                       
{'loss': -0.0561, 'grad_norm': 4.304089069366455, 'learning_rate': 6.866666666666667e-06, 'num_tokens': 2095503.0, 'completions/mean_length': 160.75, 'completions/min_length': 93.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 110.5714340209961, 'completions/min_terminated_length': 93.0, 'completions/max_terminated_length': 128.0, 'rewards/meter/mean': 0.645095944404602, 'rewards/meter/std': 0.45269909501075745, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.17251639068126678, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.958979606628418, 'rewards/lexical_plausibility/std': 0.07877527922391891, 'rewards/judge_quality/mean': 0.13750000298023224, 'rewards/judge_quality/std': 0.0353553406894207, 'rewards/repeat_penalty/mean': 0.32624149322509766, 'rewards/repeat_penalty/std': 0.22708722949028015, 'rewards/near_duplicate_penalty/mean': 0.6438634395599365, 'rewards/near_duplicate_penalty/std': 0.17047125101089478, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1530931144952774, 'rewards/distinct_2/mean': 0.49744752049446106, 'rewards/distinct_2/std': 0.27945196628570557, 'rewards/total_composite/mean': 0.07263710349798203, 'rewards/total_composite/std': 0.07002465426921844, 'reward': 0.07263710349798203, 'reward_std': 0.07002464681863785, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0699535682797432, 'sampling/sampling_logp_difference/max': 1.0906991958618164, 'sampling/importance_sampling_ratio/min': 0.33598148822784424, 'sampling/importance_sampling_ratio/mean': 1.00955331325531, 'sampling/importance_sampling_ratio/max': 1.9508916139602661, 'entropy': 0.42657043412327766, 'clip_ratio/low_mean': 0.017774964682757854, 'clip_ratio/low_min': 0.017774964682757854, 'clip_ratio/high_mean': 0.031890815356746316, 'clip_ratio/high_max': 0.031890815356746316, 'clip_ratio/region_mean': 0.04966578003950417, 'reward_total_mean': 0.07263710349798203, 'reward_meter_mean': 0.645095944404602, 'reward_meter_std': 0.45269909501075745, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.17251639068126678, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.958979606628418, 'reward_lexical_plausibility_std': 0.07877527922391891, 'reward_repeat_penalty_mean': 0.32624149322509766, 'reward_repeat_penalty_std': 0.22708722949028015, 'reward_near_duplicate_penalty_mean': 0.6438634395599365, 'reward_near_duplicate_penalty_std': 0.17047125101089478, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1530931144952774, 'reward_distinct_2_mean': 0.49744752049446106, 'reward_distinct_2_std': 0.27945196628570557, 'reward_judge_quality_mean': 0.13750000298023224, 'reward_judge_quality_std': 0.0353553406894207, 'reward_total_composite_mean': 0.07263710349798203, 'reward_total_composite_std': 0.07002465426921844, 'epoch': 0.04}

 31%|███▏      | 1035/3300 [4:09:54<7:47:51, 12.39s/it]INFO 04-12 14:56:06 [block_pool.py:378] Successfully reset prefix cache

 31%|███▏      | 1036/3300 [4:10:01<6:51:04, 10.89s/it]2026-04-12 14:56:13,746 | INFO | train_grpo_train | metrics_logged mode=train step=1036

                                                       
{'loss': -0.0272, 'grad_norm': 16.272716522216797, 'learning_rate': 6.8636363636363645e-06, 'num_tokens': 2096949.0, 'completions/mean_length': 23.75, 'completions/min_length': 21.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.75, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.29522714018821716, 'rewards/meter/std': 0.4176274538040161, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9713541269302368, 'rewards/lexical_plausibility/std': 0.05990393087267876, 'rewards/judge_quality/mean': 0.8525000214576721, 'rewards/judge_quality/std': 0.28469279408454895, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.27898114919662476, 'rewards/total_composite/std': 0.3970460295677185, 'reward': 0.27898114919662476, 'reward_std': 0.3970460295677185, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1501610428094864, 'sampling/sampling_logp_difference/max': 2.67525315284729, 'sampling/importance_sampling_ratio/min': 0.06888938695192337, 'sampling/importance_sampling_ratio/mean': 1.004130244255066, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6254268996417522, 'clip_ratio/low_mean': 0.08184818085283041, 'clip_ratio/low_min': 0.08184818085283041, 'clip_ratio/high_mean': 0.030448718927800655, 'clip_ratio/high_max': 0.030448718927800655, 'clip_ratio/region_mean': 0.11229689978063107, 'reward_total_mean': 0.27898114919662476, 'reward_meter_mean': 0.29522714018821716, 'reward_meter_std': 0.4176274538040161, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9713541269302368, 'reward_lexical_plausibility_std': 0.05990393087267876, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8525000214576721, 'reward_judge_quality_std': 0.28469279408454895, 'reward_total_composite_mean': 0.27898114919662476, 'reward_total_composite_std': 0.3970460295677185, 'epoch': 0.04}

 31%|███▏      | 1036/3300 [4:10:01<6:51:04, 10.89s/it]INFO 04-12 14:56:13 [block_pool.py:378] Successfully reset prefix cache

 31%|███▏      | 1037/3300 [4:10:10<6:27:13, 10.27s/it]2026-04-12 14:56:22,561 | INFO | train_grpo_train | metrics_logged mode=train step=1037

                                                       
{'loss': 0.0198, 'grad_norm': 6.293120384216309, 'learning_rate': 6.860606060606061e-06, 'num_tokens': 2099287.0, 'completions/mean_length': 107.25, 'completions/min_length': 95.0, 'completions/max_length': 116.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 107.25, 'completions/min_terminated_length': 95.0, 'completions/max_terminated_length': 116.0, 'rewards/meter/mean': 0.6889856457710266, 'rewards/meter/std': 0.37136074900627136, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9800558090209961, 'rewards/lexical_plausibility/std': 0.041558220982551575, 'rewards/judge_quality/mean': 0.20000000298023224, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.07567217946052551, 'rewards/repeat_penalty/std': 0.10801900178194046, 'rewards/near_duplicate_penalty/mean': 0.5517589449882507, 'rewards/near_duplicate_penalty/std': 0.19355860352516174, 'rewards/opening_diversity/mean': 0.3984375, 'rewards/opening_diversity/std': 0.1858183592557907, 'rewards/distinct_2/mean': 0.25030049681663513, 'rewards/distinct_2/std': 0.2290385514497757, 'rewards/total_composite/mean': 0.12731152772903442, 'rewards/total_composite/std': 0.08997271209955215, 'reward': 0.12731152772903442, 'reward_std': 0.08997270464897156, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07115565985441208, 'sampling/sampling_logp_difference/max': 1.5724859237670898, 'sampling/importance_sampling_ratio/min': 0.20752863585948944, 'sampling/importance_sampling_ratio/mean': 0.9952552318572998, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3931223377585411, 'clip_ratio/low_mean': 0.028525335714221, 'clip_ratio/low_min': 0.028525335714221, 'clip_ratio/high_mean': 0.03271407098509371, 'clip_ratio/high_max': 0.03271407098509371, 'clip_ratio/region_mean': 0.061239406699314713, 'reward_total_mean': 0.12731152772903442, 'reward_meter_mean': 0.6889856457710266, 'reward_meter_std': 0.37136074900627136, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9800558090209961, 'reward_lexical_plausibility_std': 0.041558220982551575, 'reward_repeat_penalty_mean': 0.07567217946052551, 'reward_repeat_penalty_std': 0.10801900178194046, 'reward_near_duplicate_penalty_mean': 0.5517589449882507, 'reward_near_duplicate_penalty_std': 0.19355860352516174, 'reward_opening_diversity_mean': 0.3984375, 'reward_opening_diversity_std': 0.1858183592557907, 'reward_distinct_2_mean': 0.25030049681663513, 'reward_distinct_2_std': 0.2290385514497757, 'reward_judge_quality_mean': 0.20000000298023224, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.12731152772903442, 'reward_total_composite_std': 0.08997271209955215, 'epoch': 0.04}

 31%|███▏      | 1037/3300 [4:10:10<6:27:13, 10.27s/it]INFO 04-12 14:56:22 [block_pool.py:378] Successfully reset prefix cache

 31%|███▏      | 1038/3300 [4:10:24<7:05:55, 11.30s/it]2026-04-12 14:56:36,257 | INFO | train_grpo_train | metrics_logged mode=train step=1038

                                                       
{'loss': 0.0065, 'grad_norm': 3.965681791305542, 'learning_rate': 6.857575757575758e-06, 'num_tokens': 2100685.0, 'completions/mean_length': 83.75, 'completions/min_length': 19.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 22.571430206298828, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.25796061754226685, 'rewards/meter/std': 0.4418087601661682, '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/lexical_plausibility/mean': 0.9571786522865295, 'rewards/lexical_plausibility/std': 0.12111710757017136, 'rewards/judge_quality/mean': 0.476250022649765, 'rewards/judge_quality/std': 0.38119879364967346, 'rewards/repeat_penalty/mean': 0.7451828122138977, 'rewards/repeat_penalty/std': 0.013625085353851318, 'rewards/near_duplicate_penalty/mean': 0.9935770630836487, 'rewards/near_duplicate_penalty/std': 0.018166767433285713, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.06684528291225433, 'rewards/total_composite/std': 0.12137099355459213, 'reward': 0.06684528291225433, 'reward_std': 0.12137098610401154, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13378006219863892, 'sampling/sampling_logp_difference/max': 2.5705623626708984, 'sampling/importance_sampling_ratio/min': 0.07649251818656921, 'sampling/importance_sampling_ratio/mean': 0.9982591867446899, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.48402202129364014, 'clip_ratio/low_mean': 0.04154491424560547, 'clip_ratio/low_min': 0.04154491424560547, 'clip_ratio/high_mean': 0.01883971318602562, 'clip_ratio/high_max': 0.01883971318602562, 'clip_ratio/region_mean': 0.06038462743163109, 'reward_total_mean': 0.06684528291225433, 'reward_meter_mean': 0.25796061754226685, 'reward_meter_std': 0.4418087601661682, '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_lexical_plausibility_mean': 0.9571786522865295, 'reward_lexical_plausibility_std': 0.12111710757017136, 'reward_repeat_penalty_mean': 0.7451828122138977, 'reward_repeat_penalty_std': 0.013625085353851318, 'reward_near_duplicate_penalty_mean': 0.9935770630836487, 'reward_near_duplicate_penalty_std': 0.018166767433285713, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.476250022649765, 'reward_judge_quality_std': 0.38119879364967346, 'reward_total_composite_mean': 0.06684528291225433, 'reward_total_composite_std': 0.12137099355459213, 'epoch': 0.04}

 31%|███▏      | 1038/3300 [4:10:24<7:05:55, 11.30s/it]INFO 04-12 14:56:36 [block_pool.py:378] Successfully reset prefix cache

 31%|███▏      | 1039/3300 [4:10:33<6:43:10, 10.70s/it]2026-04-12 14:56:45,557 | INFO | train_grpo_train | metrics_logged mode=train step=1039

                                                       
{'loss': 0.0758, 'grad_norm': 11.13637924194336, 'learning_rate': 6.854545454545455e-06, 'num_tokens': 2102432.0, 'completions/mean_length': 50.375, 'completions/min_length': 35.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.375, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.4992498755455017, 'rewards/meter/std': 0.38977953791618347, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9713541269302368, 'rewards/lexical_plausibility/std': 0.05990393087267876, 'rewards/judge_quality/mean': 0.4925000071525574, 'rewards/judge_quality/std': 0.2796298861503601, 'rewards/repeat_penalty/mean': 0.7816746234893799, 'rewards/repeat_penalty/std': 0.19714240729808807, 'rewards/near_duplicate_penalty/mean': 0.8836131691932678, 'rewards/near_duplicate_penalty/std': 0.08969806879758835, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.8982267379760742, 'rewards/distinct_2/std': 0.10319311171770096, 'rewards/total_composite/mean': 0.22295686602592468, 'rewards/total_composite/std': 0.1913090944290161, 'reward': 0.22295686602592468, 'reward_std': 0.19130907952785492, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13055534660816193, 'sampling/sampling_logp_difference/max': 2.1627912521362305, 'sampling/importance_sampling_ratio/min': 0.11500366032123566, 'sampling/importance_sampling_ratio/mean': 1.0089117288589478, 'sampling/importance_sampling_ratio/max': 1.9326164722442627, 'entropy': 0.688446007668972, 'clip_ratio/low_mean': 0.039414272643625736, 'clip_ratio/low_min': 0.039414272643625736, 'clip_ratio/high_mean': 0.059787140460684896, 'clip_ratio/high_max': 0.059787140460684896, 'clip_ratio/region_mean': 0.09920141310431063, 'reward_total_mean': 0.22295686602592468, 'reward_meter_mean': 0.4992498755455017, 'reward_meter_std': 0.38977953791618347, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9713541269302368, 'reward_lexical_plausibility_std': 0.05990393087267876, 'reward_repeat_penalty_mean': 0.7816746234893799, 'reward_repeat_penalty_std': 0.19714240729808807, 'reward_near_duplicate_penalty_mean': 0.8836131691932678, 'reward_near_duplicate_penalty_std': 0.08969806879758835, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.8982267379760742, 'reward_distinct_2_std': 0.10319311171770096, 'reward_judge_quality_mean': 0.4925000071525574, 'reward_judge_quality_std': 0.2796298861503601, 'reward_total_composite_mean': 0.22295686602592468, 'reward_total_composite_std': 0.1913090944290161, 'epoch': 0.04}

 31%|███▏      | 1039/3300 [4:10:33<6:43:10, 10.70s/it]INFO 04-12 14:56:45 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1040/3300 [4:10:42<6:28:07, 10.30s/it]2026-04-12 14:56:54,940 | INFO | train_grpo_train | metrics_logged mode=train step=1040

                                                       
{'loss': 0.0598, 'grad_norm': 9.161330223083496, 'learning_rate': 6.851515151515153e-06, 'num_tokens': 2104675.0, 'completions/mean_length': 98.375, 'completions/min_length': 86.0, 'completions/max_length': 134.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.375, 'completions/min_terminated_length': 86.0, 'completions/max_terminated_length': 134.0, 'rewards/meter/mean': 0.6199376583099365, 'rewards/meter/std': 0.29622992873191833, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.20000000298023224, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.17406870424747467, 'rewards/repeat_penalty/std': 0.14580193161964417, 'rewards/near_duplicate_penalty/mean': 0.7089067697525024, 'rewards/near_duplicate_penalty/std': 0.165482297539711, 'rewards/opening_diversity/mean': 0.6491477489471436, 'rewards/opening_diversity/std': 0.2793881595134735, 'rewards/distinct_2/mean': 0.34906089305877686, 'rewards/distinct_2/std': 0.25997501611709595, 'rewards/total_composite/mean': 0.11755747348070145, 'rewards/total_composite/std': 0.06578730046749115, 'reward': 0.11755747348070145, 'reward_std': 0.06578729301691055, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09282664209604263, 'sampling/sampling_logp_difference/max': 2.5501863956451416, 'sampling/importance_sampling_ratio/min': 0.0780671164393425, 'sampling/importance_sampling_ratio/mean': 0.9970666766166687, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.34638754837214947, 'clip_ratio/low_mean': 0.026927629485726357, 'clip_ratio/low_min': 0.026927629485726357, 'clip_ratio/high_mean': 0.05801033042371273, 'clip_ratio/high_max': 0.05801033042371273, 'clip_ratio/region_mean': 0.08493795990943909, 'reward_total_mean': 0.11755747348070145, 'reward_meter_mean': 0.6199376583099365, 'reward_meter_std': 0.29622992873191833, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.17406870424747467, 'reward_repeat_penalty_std': 0.14580193161964417, 'reward_near_duplicate_penalty_mean': 0.7089067697525024, 'reward_near_duplicate_penalty_std': 0.165482297539711, 'reward_opening_diversity_mean': 0.6491477489471436, 'reward_opening_diversity_std': 0.2793881595134735, 'reward_distinct_2_mean': 0.34906089305877686, 'reward_distinct_2_std': 0.25997501611709595, 'reward_judge_quality_mean': 0.20000000298023224, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.11755747348070145, 'reward_total_composite_std': 0.06578730046749115, 'epoch': 0.04}

 32%|███▏      | 1040/3300 [4:10:42<6:28:07, 10.30s/it]INFO 04-12 14:56:55 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1041/3300 [4:10:55<6:59:37, 11.15s/it]2026-04-12 14:57:08,046 | INFO | train_grpo_train | metrics_logged mode=train step=1041

                                                       
{'loss': -0.0489, 'grad_norm': 3.3286194801330566, 'learning_rate': 6.848484848484849e-06, 'num_tokens': 2106082.0, 'completions/mean_length': 85.875, 'completions/min_length': 12.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 25.000001907348633, 'completions/min_terminated_length': 12.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.5926029682159424, 'rewards/meter/std': 0.4901561141014099, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.8277760148048401, 'rewards/lexical_plausibility/std': 0.3452592194080353, 'rewards/judge_quality/mean': 0.6637499928474426, 'rewards/judge_quality/std': 0.3867792785167694, 'rewards/repeat_penalty/mean': 0.6132802367210388, 'rewards/repeat_penalty/std': 0.28017112612724304, 'rewards/near_duplicate_penalty/mean': 0.9526224136352539, 'rewards/near_duplicate_penalty/std': 0.06800727546215057, 'rewards/opening_diversity/mean': 0.6875, 'rewards/opening_diversity/std': 0.29124119877815247, 'rewards/distinct_2/mean': 0.9132781028747559, 'rewards/distinct_2/std': 0.07122636586427689, 'rewards/total_composite/mean': 0.49210023880004883, 'rewards/total_composite/std': 0.43593814969062805, 'reward': 0.49210023880004883, 'reward_std': 0.43593811988830566, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10442553460597992, 'sampling/sampling_logp_difference/max': 1.0248875617980957, 'sampling/importance_sampling_ratio/min': 0.358836829662323, 'sampling/importance_sampling_ratio/mean': 0.9949103593826294, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.41358262300491333, 'clip_ratio/low_mean': 0.020717463456094265, 'clip_ratio/low_min': 0.020717463456094265, 'clip_ratio/high_mean': 0.029761905316263437, 'clip_ratio/high_max': 0.029761905316263437, 'clip_ratio/region_mean': 0.0504793687723577, 'reward_total_mean': 0.49210023880004883, 'reward_meter_mean': 0.5926029682159424, 'reward_meter_std': 0.4901561141014099, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.8277760148048401, 'reward_lexical_plausibility_std': 0.3452592194080353, 'reward_repeat_penalty_mean': 0.6132802367210388, 'reward_repeat_penalty_std': 0.28017112612724304, 'reward_near_duplicate_penalty_mean': 0.9526224136352539, 'reward_near_duplicate_penalty_std': 0.06800727546215057, 'reward_opening_diversity_mean': 0.6875, 'reward_opening_diversity_std': 0.29124119877815247, 'reward_distinct_2_mean': 0.9132781028747559, 'reward_distinct_2_std': 0.07122636586427689, 'reward_judge_quality_mean': 0.6637499928474426, 'reward_judge_quality_std': 0.3867792785167694, 'reward_total_composite_mean': 0.49210023880004883, 'reward_total_composite_std': 0.43593814969062805, 'epoch': 0.04}

 32%|███▏      | 1041/3300 [4:10:56<6:59:37, 11.15s/it]INFO 04-12 14:57:08 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1042/3300 [4:11:06<6:50:03, 10.90s/it]2026-04-12 14:57:18,361 | INFO | train_grpo_train | metrics_logged mode=train step=1042

                                                       
{'loss': 0.0202, 'grad_norm': 5.643406391143799, 'learning_rate': 6.845454545454546e-06, 'num_tokens': 2108443.0, 'completions/mean_length': 124.125, 'completions/min_length': 118.0, 'completions/max_length': 136.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 124.125, 'completions/min_terminated_length': 118.0, 'completions/max_terminated_length': 136.0, 'rewards/meter/mean': 0.3518877625465393, 'rewards/meter/std': 0.40807414054870605, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.18866121768951416, 'rewards/repeat_penalty/std': 0.1627783626317978, 'rewards/near_duplicate_penalty/mean': 0.6615521907806396, 'rewards/near_duplicate_penalty/std': 0.10860022902488708, 'rewards/opening_diversity/mean': 0.8671875, 'rewards/opening_diversity/std': 0.21505165100097656, 'rewards/distinct_2/mean': 0.33811962604522705, 'rewards/distinct_2/std': 0.20475278794765472, 'rewards/total_composite/mean': 0.052783168852329254, 'rewards/total_composite/std': 0.061211127787828445, 'reward': 0.052783168852329254, 'reward_std': 0.061211127787828445, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06198493763804436, 'sampling/sampling_logp_difference/max': 2.6736860275268555, 'sampling/importance_sampling_ratio/min': 0.06899742782115936, 'sampling/importance_sampling_ratio/mean': 1.0036184787750244, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.39221990294754505, 'clip_ratio/low_mean': 0.027916133869439363, 'clip_ratio/low_min': 0.027916133869439363, 'clip_ratio/high_mean': 0.032405747566372156, 'clip_ratio/high_max': 0.032405747566372156, 'clip_ratio/region_mean': 0.06032188143581152, 'reward_total_mean': 0.052783168852329254, 'reward_meter_mean': 0.3518877625465393, 'reward_meter_std': 0.40807414054870605, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.18866121768951416, 'reward_repeat_penalty_std': 0.1627783626317978, 'reward_near_duplicate_penalty_mean': 0.6615521907806396, 'reward_near_duplicate_penalty_std': 0.10860022902488708, 'reward_opening_diversity_mean': 0.8671875, 'reward_opening_diversity_std': 0.21505165100097656, 'reward_distinct_2_mean': 0.33811962604522705, 'reward_distinct_2_std': 0.20475278794765472, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.052783168852329254, 'reward_total_composite_std': 0.061211127787828445, 'epoch': 0.04}

 32%|███▏      | 1042/3300 [4:11:06<6:50:03, 10.90s/it]INFO 04-12 14:57:18 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1043/3300 [4:11:15<6:28:47, 10.34s/it]2026-04-12 14:57:27,400 | INFO | train_grpo_train | metrics_logged mode=train step=1043

                                                       
{'loss': -0.0035, 'grad_norm': 11.984976768493652, 'learning_rate': 6.842424242424243e-06, 'num_tokens': 2110083.0, 'completions/mean_length': 50.0, 'completions/min_length': 43.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.0, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.497693806886673, 'rewards/meter/std': 0.4275621771812439, '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/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.08864051848649979, 'rewards/repeat_penalty/mean': 0.7967106103897095, 'rewards/repeat_penalty/std': 0.12067160755395889, 'rewards/near_duplicate_penalty/mean': 0.8925138711929321, 'rewards/near_duplicate_penalty/std': 0.058820899575948715, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8885694146156311, 'rewards/distinct_2/std': 0.07826769351959229, 'rewards/total_composite/mean': 0.14110656082630157, 'rewards/total_composite/std': 0.14107048511505127, 'reward': 0.14110656082630157, 'reward_std': 0.14107047021389008, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11833526194095612, 'sampling/sampling_logp_difference/max': 1.821230173110962, 'sampling/importance_sampling_ratio/min': 0.16182655096054077, 'sampling/importance_sampling_ratio/mean': 1.0077437162399292, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6130513288080692, 'clip_ratio/low_mean': 0.06046593235805631, 'clip_ratio/low_min': 0.06046593235805631, 'clip_ratio/high_mean': 0.031651423778384924, 'clip_ratio/high_max': 0.031651423778384924, 'clip_ratio/region_mean': 0.09211735613644123, 'reward_total_mean': 0.14110656082630157, 'reward_meter_mean': 0.497693806886673, 'reward_meter_std': 0.4275621771812439, '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_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7967106103897095, 'reward_repeat_penalty_std': 0.12067160755395889, 'reward_near_duplicate_penalty_mean': 0.8925138711929321, 'reward_near_duplicate_penalty_std': 0.058820899575948715, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8885694146156311, 'reward_distinct_2_std': 0.07826769351959229, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.08864051848649979, 'reward_total_composite_mean': 0.14110656082630157, 'reward_total_composite_std': 0.14107048511505127, 'epoch': 0.04}

 32%|███▏      | 1043/3300 [4:11:15<6:28:47, 10.34s/it]INFO 04-12 14:57:27 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1044/3300 [4:11:24<6:15:08,  9.98s/it]2026-04-12 14:57:36,531 | INFO | train_grpo_train | metrics_logged mode=train step=1044

                                                       
{'loss': 0.0425, 'grad_norm': 9.679908752441406, 'learning_rate': 6.83939393939394e-06, 'num_tokens': 2111976.0, 'completions/mean_length': 58.625, 'completions/min_length': 49.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.625, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.953931450843811, 'rewards/meter/std': 0.03660731762647629, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.23750001192092896, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.2892645001411438, 'rewards/repeat_penalty/std': 0.18122747540473938, 'rewards/near_duplicate_penalty/mean': 0.6895205974578857, 'rewards/near_duplicate_penalty/std': 0.17486591637134552, 'rewards/opening_diversity/mean': 0.5625, 'rewards/opening_diversity/std': 0.20044593513011932, 'rewards/distinct_2/mean': 0.6810886859893799, 'rewards/distinct_2/std': 0.17342659831047058, 'rewards/total_composite/mean': 0.22390568256378174, 'rewards/total_composite/std': 0.11150608211755753, 'reward': 0.22390568256378174, 'reward_std': 0.11150607466697693, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08960143476724625, 'sampling/sampling_logp_difference/max': 1.7579872608184814, 'sampling/importance_sampling_ratio/min': 0.17239150404930115, 'sampling/importance_sampling_ratio/mean': 1.006904125213623, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4613865576684475, 'clip_ratio/low_mean': 0.03758245659992099, 'clip_ratio/low_min': 0.03758245659992099, 'clip_ratio/high_mean': 0.03716727206483483, 'clip_ratio/high_max': 0.03716727206483483, 'clip_ratio/region_mean': 0.07474972866475582, 'reward_total_mean': 0.22390568256378174, 'reward_meter_mean': 0.953931450843811, 'reward_meter_std': 0.03660731762647629, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.2892645001411438, 'reward_repeat_penalty_std': 0.18122747540473938, 'reward_near_duplicate_penalty_mean': 0.6895205974578857, 'reward_near_duplicate_penalty_std': 0.17486591637134552, 'reward_opening_diversity_mean': 0.5625, 'reward_opening_diversity_std': 0.20044593513011932, 'reward_distinct_2_mean': 0.6810886859893799, 'reward_distinct_2_std': 0.17342659831047058, 'reward_judge_quality_mean': 0.23750001192092896, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.22390568256378174, 'reward_total_composite_std': 0.11150608211755753, 'epoch': 0.04}

 32%|███▏      | 1044/3300 [4:11:24<6:15:08,  9.98s/it]INFO 04-12 14:57:36 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1045/3300 [4:11:32<5:54:42,  9.44s/it]2026-04-12 14:57:44,723 | INFO | train_grpo_train | metrics_logged mode=train step=1045

                                                       
{'loss': -0.0216, 'grad_norm': 8.130537033081055, 'learning_rate': 6.8363636363636364e-06, 'num_tokens': 2113740.0, 'completions/mean_length': 58.5, 'completions/min_length': 53.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.5, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9601110219955444, 'rewards/meter/std': 0.05813397839665413, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.07559289038181305, 'rewards/repeat_penalty/mean': 0.16896995902061462, 'rewards/repeat_penalty/std': 0.10778757929801941, 'rewards/near_duplicate_penalty/mean': 0.7068864107131958, 'rewards/near_duplicate_penalty/std': 0.16768382489681244, 'rewards/opening_diversity/mean': 0.375, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.6246208548545837, 'rewards/distinct_2/std': 0.25567397475242615, 'rewards/total_composite/mean': 0.28806552290916443, 'rewards/total_composite/std': 0.07587701082229614, 'reward': 0.28806552290916443, 'reward_std': 0.07587700337171555, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08639981597661972, 'sampling/sampling_logp_difference/max': 1.9444198608398438, 'sampling/importance_sampling_ratio/min': 0.14307019114494324, 'sampling/importance_sampling_ratio/mean': 0.9960545301437378, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.41695987060666084, 'clip_ratio/low_mean': 0.027110593393445015, 'clip_ratio/low_min': 0.027110593393445015, 'clip_ratio/high_mean': 0.04287410504184663, 'clip_ratio/high_max': 0.04287410504184663, 'clip_ratio/region_mean': 0.06998469843529165, 'reward_total_mean': 0.28806552290916443, 'reward_meter_mean': 0.9601110219955444, 'reward_meter_std': 0.05813397839665413, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.16896995902061462, 'reward_repeat_penalty_std': 0.10778757929801941, 'reward_near_duplicate_penalty_mean': 0.7068864107131958, 'reward_near_duplicate_penalty_std': 0.16768382489681244, 'reward_opening_diversity_mean': 0.375, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.6246208548545837, 'reward_distinct_2_std': 0.25567397475242615, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.07559289038181305, 'reward_total_composite_mean': 0.28806552290916443, 'reward_total_composite_std': 0.07587701082229614, 'epoch': 0.04}

 32%|███▏      | 1045/3300 [4:11:32<5:54:42,  9.44s/it]INFO 04-12 14:57:44 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1046/3300 [4:11:41<5:46:32,  9.22s/it]2026-04-12 14:57:53,445 | INFO | train_grpo_train | metrics_logged mode=train step=1046

                                                       
{'loss': 0.0385, 'grad_norm': 12.937902450561523, 'learning_rate': 6.833333333333334e-06, 'num_tokens': 2115293.0, 'completions/mean_length': 51.125, 'completions/min_length': 45.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.125, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.16545015573501587, 'rewards/meter/std': 0.10467123240232468, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.8051742315292358, 'rewards/repeat_penalty/std': 0.19072328507900238, 'rewards/near_duplicate_penalty/mean': 0.9040540456771851, 'rewards/near_duplicate_penalty/std': 0.060803644359111786, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9030629396438599, 'rewards/distinct_2/std': 0.11608945578336716, 'rewards/total_composite/mean': 0.054583821445703506, 'rewards/total_composite/std': 0.028577309101819992, 'reward': 0.054583821445703506, 'reward_std': 0.02857731096446514, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12414848059415817, 'sampling/sampling_logp_difference/max': 1.4196252822875977, 'sampling/importance_sampling_ratio/min': 0.2418045997619629, 'sampling/importance_sampling_ratio/mean': 0.9813938736915588, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5275466926395893, 'clip_ratio/low_mean': 0.048214022535830736, 'clip_ratio/low_min': 0.048214022535830736, 'clip_ratio/high_mean': 0.05502551142126322, 'clip_ratio/high_max': 0.05502551142126322, 'clip_ratio/region_mean': 0.10323953395709395, 'reward_total_mean': 0.054583821445703506, 'reward_meter_mean': 0.16545015573501587, 'reward_meter_std': 0.10467123240232468, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8051742315292358, 'reward_repeat_penalty_std': 0.19072328507900238, 'reward_near_duplicate_penalty_mean': 0.9040540456771851, 'reward_near_duplicate_penalty_std': 0.060803644359111786, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9030629396438599, 'reward_distinct_2_std': 0.11608945578336716, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.054583821445703506, 'reward_total_composite_std': 0.028577309101819992, 'epoch': 0.04}

 32%|███▏      | 1046/3300 [4:11:41<5:46:32,  9.22s/it]INFO 04-12 14:57:53 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1047/3300 [4:11:55<6:37:27, 10.58s/it]2026-04-12 14:58:07,197 | INFO | train_grpo_train | metrics_logged mode=train step=1047

                                                       
{'loss': -0.1305, 'grad_norm': 2.625126600265503, 'learning_rate': 6.83030303030303e-06, 'num_tokens': 2117785.0, 'completions/mean_length': 199.5, 'completions/min_length': 143.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 154.85714721679688, 'completions/min_terminated_length': 143.0, 'completions/max_terminated_length': 190.0, 'rewards/meter/mean': 0.6465119123458862, 'rewards/meter/std': 0.47260215878486633, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.1178511381149292, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/lexical_plausibility/mean': 0.9783613681793213, 'rewards/lexical_plausibility/std': 0.06120318919420242, 'rewards/judge_quality/mean': 0.13750000298023224, 'rewards/judge_quality/std': 0.0353553406894207, 'rewards/repeat_penalty/mean': 0.11797183752059937, 'rewards/repeat_penalty/std': 0.25789815187454224, 'rewards/near_duplicate_penalty/mean': 0.4955127239227295, 'rewards/near_duplicate_penalty/std': 0.23408250510692596, 'rewards/opening_diversity/mean': 0.484375, 'rewards/opening_diversity/std': 0.16952534019947052, 'rewards/distinct_2/mean': 0.22648878395557404, 'rewards/distinct_2/std': 0.3366652727127075, 'rewards/total_composite/mean': 0.0628373920917511, 'rewards/total_composite/std': 0.049676164984703064, 'reward': 0.0628373920917511, 'reward_std': 0.04967616870999336, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.043361932039260864, 'sampling/sampling_logp_difference/max': 1.4238332509994507, 'sampling/importance_sampling_ratio/min': 0.2593633532524109, 'sampling/importance_sampling_ratio/mean': 1.0077636241912842, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2667356953024864, 'clip_ratio/low_mean': 0.01226362632587552, 'clip_ratio/low_min': 0.01226362632587552, 'clip_ratio/high_mean': 0.022837560391053557, 'clip_ratio/high_max': 0.022837560391053557, 'clip_ratio/region_mean': 0.03510118671692908, 'reward_total_mean': 0.0628373920917511, 'reward_meter_mean': 0.6465119123458862, 'reward_meter_std': 0.47260215878486633, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.1178511381149292, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_lexical_plausibility_mean': 0.9783613681793213, 'reward_lexical_plausibility_std': 0.06120318919420242, 'reward_repeat_penalty_mean': 0.11797183752059937, 'reward_repeat_penalty_std': 0.25789815187454224, 'reward_near_duplicate_penalty_mean': 0.4955127239227295, 'reward_near_duplicate_penalty_std': 0.23408250510692596, 'reward_opening_diversity_mean': 0.484375, 'reward_opening_diversity_std': 0.16952534019947052, 'reward_distinct_2_mean': 0.22648878395557404, 'reward_distinct_2_std': 0.3366652727127075, 'reward_judge_quality_mean': 0.13750000298023224, 'reward_judge_quality_std': 0.0353553406894207, 'reward_total_composite_mean': 0.0628373920917511, 'reward_total_composite_std': 0.049676164984703064, 'epoch': 0.04}

 32%|███▏      | 1047/3300 [4:11:55<6:37:27, 10.58s/it]INFO 04-12 14:58:07 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1048/3300 [4:12:03<6:17:09, 10.05s/it]2026-04-12 14:58:16,080 | INFO | train_grpo_train | metrics_logged mode=train step=1048

                                                       
{'loss': 0.02, 'grad_norm': 17.605117797851562, 'learning_rate': 6.827272727272728e-06, 'num_tokens': 2119243.0, 'completions/mean_length': 31.25, 'completions/min_length': 28.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.25, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.895077109336853, 'rewards/meter/std': 0.27064019441604614, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9173212051391602, 'rewards/lexical_plausibility/std': 0.209740549325943, 'rewards/judge_quality/mean': 0.4424999952316284, 'rewards/judge_quality/std': 0.32101404666900635, 'rewards/repeat_penalty/mean': 0.640625, 'rewards/repeat_penalty/std': 0.26252126693725586, 'rewards/near_duplicate_penalty/mean': 0.9501262903213501, 'rewards/near_duplicate_penalty/std': 0.09400057047605515, 'rewards/opening_diversity/mean': 0.65625, 'rewards/opening_diversity/std': 0.2651650309562683, 'rewards/distinct_2/mean': 0.9807692170143127, 'rewards/distinct_2/std': 0.054392825812101364, 'rewards/total_composite/mean': 0.3957640826702118, 'rewards/total_composite/std': 0.34027931094169617, 'reward': 0.3957640826702118, 'reward_std': 0.34027931094169617, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14961084723472595, 'sampling/sampling_logp_difference/max': 3.684997797012329, 'sampling/importance_sampling_ratio/min': 0.025097230449318886, 'sampling/importance_sampling_ratio/mean': 0.9999329447746277, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8068030849099159, 'clip_ratio/low_mean': 0.0725634228438139, 'clip_ratio/low_min': 0.0725634228438139, 'clip_ratio/high_mean': 0.040982948150485754, 'clip_ratio/high_max': 0.040982948150485754, 'clip_ratio/region_mean': 0.11354637099429965, 'reward_total_mean': 0.3957640826702118, 'reward_meter_mean': 0.895077109336853, 'reward_meter_std': 0.27064019441604614, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 0.9173212051391602, 'reward_lexical_plausibility_std': 0.209740549325943, 'reward_repeat_penalty_mean': 0.640625, 'reward_repeat_penalty_std': 0.26252126693725586, 'reward_near_duplicate_penalty_mean': 0.9501262903213501, 'reward_near_duplicate_penalty_std': 0.09400057047605515, 'reward_opening_diversity_mean': 0.65625, 'reward_opening_diversity_std': 0.2651650309562683, 'reward_distinct_2_mean': 0.9807692170143127, 'reward_distinct_2_std': 0.054392825812101364, 'reward_judge_quality_mean': 0.4424999952316284, 'reward_judge_quality_std': 0.32101404666900635, 'reward_total_composite_mean': 0.3957640826702118, 'reward_total_composite_std': 0.34027931094169617, 'epoch': 0.04}

 32%|███▏      | 1048/3300 [4:12:04<6:17:09, 10.05s/it]INFO 04-12 14:58:16 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1049/3300 [4:12:13<6:09:50,  9.86s/it]2026-04-12 14:58:25,411 | INFO | train_grpo_train | metrics_logged mode=train step=1049

                                                       
{'loss': 0.1283, 'grad_norm': 4.576805114746094, 'learning_rate': 6.824242424242425e-06, 'num_tokens': 2121287.0, 'completions/mean_length': 97.5, 'completions/min_length': 80.0, 'completions/max_length': 131.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 97.5, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 131.0, 'rewards/meter/mean': 0.8802154064178467, 'rewards/meter/std': 0.16540750861167908, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.17251639068126678, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.0048208110965788364, 'rewards/repeat_penalty/std': 0.007432539016008377, 'rewards/near_duplicate_penalty/mean': 0.3683338165283203, 'rewards/near_duplicate_penalty/std': 0.06853021681308746, 'rewards/opening_diversity/mean': 0.2265625, 'rewards/opening_diversity/std': 0.032346826046705246, 'rewards/distinct_2/mean': 0.06965675950050354, 'rewards/distinct_2/std': 0.07852064073085785, 'rewards/total_composite/mean': 0.11638369411230087, 'rewards/total_composite/std': 0.033883191645145416, 'reward': 0.11638369411230087, 'reward_std': 0.033883191645145416, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04824488237500191, 'sampling/sampling_logp_difference/max': 2.618191719055176, 'sampling/importance_sampling_ratio/min': 0.07293463498353958, 'sampling/importance_sampling_ratio/mean': 1.0019862651824951, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.21380569599568844, 'clip_ratio/low_mean': 0.02046280587092042, 'clip_ratio/low_min': 0.02046280587092042, 'clip_ratio/high_mean': 0.02666075900197029, 'clip_ratio/high_max': 0.02666075900197029, 'clip_ratio/region_mean': 0.04712356487289071, 'reward_total_mean': 0.11638369411230087, 'reward_meter_mean': 0.8802154064178467, 'reward_meter_std': 0.16540750861167908, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.17251639068126678, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.0048208110965788364, 'reward_repeat_penalty_std': 0.007432539016008377, 'reward_near_duplicate_penalty_mean': 0.3683338165283203, 'reward_near_duplicate_penalty_std': 0.06853021681308746, 'reward_opening_diversity_mean': 0.2265625, 'reward_opening_diversity_std': 0.032346826046705246, 'reward_distinct_2_mean': 0.06965675950050354, 'reward_distinct_2_std': 0.07852064073085785, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.11638369411230087, 'reward_total_composite_std': 0.033883191645145416, 'epoch': 0.04}

 32%|███▏      | 1049/3300 [4:12:13<6:09:50,  9.86s/it]INFO 04-12 14:58:25 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1050/3300 [4:12:21<5:53:31,  9.43s/it]2026-04-12 14:58:33,826 | INFO | train_grpo_train | metrics_logged mode=train step=1050

                                                       
{'loss': 0.044, 'grad_norm': 9.839877128601074, 'learning_rate': 6.821212121212122e-06, 'num_tokens': 2122944.0, 'completions/mean_length': 45.125, 'completions/min_length': 39.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.125, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.28557300567626953, 'rewards/meter/std': 0.28182196617126465, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6225000023841858, 'rewards/judge_quality/std': 0.3285357356071472, 'rewards/repeat_penalty/mean': 0.7773823738098145, 'rewards/repeat_penalty/std': 0.18026597797870636, 'rewards/near_duplicate_penalty/mean': 0.9530285596847534, 'rewards/near_duplicate_penalty/std': 0.043377574533224106, 'rewards/opening_diversity/mean': 0.925000011920929, 'rewards/opening_diversity/std': 0.11338934302330017, 'rewards/distinct_2/mean': 0.8699880838394165, 'rewards/distinct_2/std': 0.096328504383564, 'rewards/total_composite/mean': 0.126708984375, 'rewards/total_composite/std': 0.09142221510410309, 'reward': 0.126708984375, 'reward_std': 0.09142221510410309, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11328752338886261, 'sampling/sampling_logp_difference/max': 1.7761284112930298, 'sampling/importance_sampling_ratio/min': 0.16929231584072113, 'sampling/importance_sampling_ratio/mean': 1.0014660358428955, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4858452193439007, 'clip_ratio/low_mean': 0.036599186481907964, 'clip_ratio/low_min': 0.036599186481907964, 'clip_ratio/high_mean': 0.050821324810385704, 'clip_ratio/high_max': 0.050821324810385704, 'clip_ratio/region_mean': 0.08742051129229367, 'reward_total_mean': 0.126708984375, 'reward_meter_mean': 0.28557300567626953, 'reward_meter_std': 0.28182196617126465, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7773823738098145, 'reward_repeat_penalty_std': 0.18026597797870636, 'reward_near_duplicate_penalty_mean': 0.9530285596847534, 'reward_near_duplicate_penalty_std': 0.043377574533224106, 'reward_opening_diversity_mean': 0.925000011920929, 'reward_opening_diversity_std': 0.11338934302330017, 'reward_distinct_2_mean': 0.8699880838394165, 'reward_distinct_2_std': 0.096328504383564, 'reward_judge_quality_mean': 0.6225000023841858, 'reward_judge_quality_std': 0.3285357356071472, 'reward_total_composite_mean': 0.126708984375, 'reward_total_composite_std': 0.09142221510410309, 'epoch': 0.04}

 32%|███▏      | 1050/3300 [4:12:21<5:53:31,  9.43s/it]INFO 04-12 14:58:34 [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 [00:10<00:59,  5.39s/it]

 23%|██▎       | 3/13 [00:17<00:59,  5.93s/it]

 31%|███       | 4/13 [00:28<01:08,  7.65s/it]

 38%|███▊      | 5/13 [00:39<01:11,  8.88s/it]

 46%|████▌     | 6/13 [00:49<01:06,  9.45s/it]

 54%|█████▍    | 7/13 [01:01<01:00, 10.06s/it]

 62%|██████▏   | 8/13 [01:12<00:51, 10.37s/it]

 69%|██████▉   | 9/13 [01:23<00:42, 10.72s/it]

 77%|███████▋  | 10/13 [01:35<00:32, 10.93s/it]

 85%|████████▍ | 11/13 [01:45<00:21, 10.81s/it]

 92%|█████████▏| 12/13 [01:56<00:10, 10.73s/it]

100%|██████████| 13/13 [02:07<00:00, 10.78s/it]2026-04-12 15:00:52,640 | INFO | train_grpo_train | metrics_logged mode=eval step=1050

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 138.7371, 'eval_samples_per_second': 0.75, 'eval_steps_per_second': 0.094, 'eval_num_tokens': 2122944.0, 'eval_completions/mean_length': 276.0673076923077, 'eval_completions/min_length': 48.53846153846154, 'eval_completions/max_length': 488.6923076923077, 'eval_completions/clipped_ratio': 0.33653846153846156, 'eval_completions/mean_terminated_length': 157.61685356727014, 'eval_completions/min_terminated_length': 48.53846153846154, 'eval_completions/max_terminated_length': 287.7692307692308, 'eval_rewards/meter/mean': 0.5317095586886773, 'eval_rewards/meter/std': 0.42014962434768677, 'eval_rewards/count_adherence/mean': 0.6783530253630418, 'eval_rewards/count_adherence/std': 0.3206257911828848, 'eval_rewards/arabic_clean/mean': 0.9615384615384616, 'eval_rewards/arabic_clean/std': 0.10878565678229699, 'eval_rewards/lexical_plausibility/mean': 0.9860358421619122, 'eval_rewards/lexical_plausibility/std': 0.03762838330406409, 'eval_rewards/judge_quality/mean': 0.16836538624304992, 'eval_rewards/judge_quality/std': 0.13044744586715332, 'eval_rewards/repeat_penalty/mean': 0.16069939271143924, 'eval_rewards/repeat_penalty/std': 0.2769986059821139, 'eval_rewards/near_duplicate_penalty/mean': 0.4656345316996941, 'eval_rewards/near_duplicate_penalty/std': 0.29570043889375835, 'eval_rewards/opening_diversity/mean': 0.5071448316940894, 'eval_rewards/opening_diversity/std': 0.3606924471946863, 'eval_rewards/distinct_2/mean': 0.2279228991112457, 'eval_rewards/distinct_2/std': 0.34475924633443356, 'eval_rewards/total_composite/mean': 0.06406389520718501, 'eval_rewards/total_composite/std': 0.08801526576280594, 'eval_reward': 0.06406389520718501, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.022444536336339437, 'eval_sampling/sampling_logp_difference/max': 0.9237296397869403, 'eval_sampling/importance_sampling_ratio/min': 0.4081879464479593, 'eval_sampling/importance_sampling_ratio/mean': 1.0031032608105586, 'eval_sampling/importance_sampling_ratio/max': 1.3501007832013643, 'eval_entropy': 0.21908611574998269, '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': 0.06406389520718501, 'eval_reward_meter_mean': 0.5317095586886773, 'eval_reward_meter_std': 0.42014962434768677, 'eval_reward_count_adherence_mean': 0.6783530253630418, 'eval_reward_count_adherence_std': 0.3206257911828848, 'eval_reward_arabic_clean_mean': 0.9615384615384616, 'eval_reward_arabic_clean_std': 0.10878565678229699, 'eval_reward_lexical_plausibility_mean': 0.9860358421619122, 'eval_reward_lexical_plausibility_std': 0.03762838330406409, 'eval_reward_repeat_penalty_mean': 0.16069939271143924, 'eval_reward_repeat_penalty_std': 0.2769986059821139, 'eval_reward_near_duplicate_penalty_mean': 0.4656345316996941, 'eval_reward_near_duplicate_penalty_std': 0.29570043889375835, 'eval_reward_opening_diversity_mean': 0.5071448316940894, 'eval_reward_opening_diversity_std': 0.3606924471946863, 'eval_reward_distinct_2_mean': 0.2279228991112457, 'eval_reward_distinct_2_std': 0.34475924633443356, 'eval_reward_judge_quality_mean': 0.16836538624304992, 'eval_reward_judge_quality_std': 0.13044744586715332, 'eval_reward_total_composite_mean': 0.06406389520718501, 'eval_reward_total_composite_std': 0.08801526576280594, 'epoch': 0.04}

 32%|███▏      | 1050/3300 [4:14:40<5:53:31,  9.43s/it]

100%|██████████| 13/13 [02:07<00:00, 10.78s/it]

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